feat(policies): add VLA-JEPA (#3568)

* first commit

* feat(policies): add VLA-JEPA

* feat(policies): add VLA-JEPA

* support vla_jepa

* (feat)policies: add VLA-JEPA

* linting

* adding deps to pyproject.toml

* updating uv lock

* adding guards to avoid needing transformers and diffusers for type checking and basic tests

* fixing action and state dim

* fix warnings with qwen processor kwargs

* fixing wm_loss not propagating

* adjusting obs steps, tublets size to match original implementation

* some more fixes to be closer to the original implem

* adding more tests to ensure good coverage

* align VLA-JEPA architecture with original checkpoint

- Remove stale `action_num_heads` / `action_attention_head_dim` config fields;
  DiT head dimensions are now always derived from the preset (DiT-B/L/test).
- Add `num_target_vision_tokens` and `action_max_seq_len` config fields required
  by the action head's future-token embedding and positional embedding tables.
- Fix default `qwen_model_name` to 2B (matches all released checkpoints).
- Rename `ActionEncoder` attrs w1/w2/w3 → layer1/layer2/layer3 to match
  checkpoint key names; replace `nn.Sequential` decoder/state-encoder with
  `_MLP2` (layer1/layer2 naming).
- Fix `VLAJEPAActionHead` to size ActionEncoder and StateEncoder at `inner_dim`
  (DiT input width) rather than `action_hidden_size` (DiT output width).
- Rename `DiT.blocks` → `transformer_blocks` and `attn` → `attn1` to match
  checkpoint; add alternating cross/self attention (even blocks cross-attend to
  Qwen context, odd blocks self-attend).
- Add `DiT-test` preset for unit tests.
- Rewrite `ActionConditionedVideoPredictor` with explicit ViT-style blocks
  (`_PredictorBlock` with fused qkv) to match checkpoint structure; rename
  `encoder`/`norm`/`proj` → `predictor_blocks`/`predictor_norm`/`predictor_proj`.

* propagate action_is_pad masking through VLA-JEPA policy pipeline

Pass the `action_is_pad` tensor from the batch through to the action head
so padded timesteps are excluded from the flow-matching loss.

* update VLA-JEPA tests for arch changes and action_is_pad

- Switch conftest to use `action_model_type="DiT-test"` now that
  `action_num_heads` / `action_attention_head_dim` have been removed.
- Add action_head tests covering fully-padded loss (zero) and equivalence
  of action_is_pad=None vs all-zeros mask.
- Remove obsolete `test_native_to_lerobot_wm_only` test.

* add VLA-JEPA documentation

Covers architecture overview, pretrained checkpoints, config reference,
training/eval commands for LIBERO-10, and guidance on fine-tuning for
single-camera datasets.

* add one-shot script to convert ginwind/VLA-JEPA checkpoints to safetensors (will remove once migrated)

* make default params more aligned with paper and pretrained models
- adding possibility of freezing qwen backbone and world model
- added tests for weight loading

* trying out to re-init the action head to avoid pretraining dimension mismatch

* allow different state dim and action dim

* removing missleading future_action_window_size to just use chunk_size

* lots of changes to make existing weights work, need to massively refactor the pre and post processing

* refactoring into using pre and post processor

* pre-commit cleanup

* fixing doc defaults args

Signed-off-by: Maxime Ellerbach <maxime@ellerbach.net>

* adressing dtype zeros issue

* adding guard for diffusers

* fixing training and exal examples

* trying to close success rate gap

* fix qwen norm layer output libero eval is now as expected

* adding instructions for different embodiement + fixing some tests

* smol fix to avoid having default CPU device when training

* fixing misconception about multiview / singleview handling

* removing conversion script

* adding licences

* adding .mdx docs and shortening polivy_vla_jepa_README.md

* removing useless pre-processor

* cleanup

* removing swish in favor of silu

* adding configuration gripper index and threshold

* fixing simlink

---------

Signed-off-by: Maxime Ellerbach <maxime@ellerbach.net>
Co-authored-by: ginwind <ginwind@mail.ustc.edu.cn>

tests/policies/vla_jepa/conftest.py

@@ -0,0 +1,273 @@
+#!/usr/bin/env python
+"""Shared fixtures and helpers for VLA-JEPA tests."""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+import torch
+from PIL import Image
+from torch import Tensor, nn
+
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
+
+# ---------------------------------------------------------------------------
+# Shared constants
+# ---------------------------------------------------------------------------
+
+BATCH_SIZE = 2
+ACTION_DIM = 3
+STATE_DIM = 4
+IMAGE_SIZE = 8
+ACTION_HORIZON = 4
+N_ACTION_STEPS = 2
+NUM_VIDEO_FRAMES = 3
+QWEN_HIDDEN_SIZE = 16  # hidden size produced by _FakeQwenBackbone
+
+EXPECTED_ACTION_CHUNK_SHAPE = (BATCH_SIZE, ACTION_HORIZON, ACTION_DIM)
+EXPECTED_SELECT_ACTION_SHAPE = (BATCH_SIZE, ACTION_DIM)
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def set_seed_all(seed: int) -> None:
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
+        torch.cuda.manual_seed_all(seed)
+
+
+def make_config(
+    action_dim: int = ACTION_DIM,
+    state_dim: int = STATE_DIM,
+    action_horizon: int = ACTION_HORIZON,
+    num_video_frames: int = NUM_VIDEO_FRAMES,
+) -> VLAJEPAConfig:
+    config = VLAJEPAConfig(
+        input_features={
+            f"{OBS_IMAGES}.laptop": PolicyFeature(type=FeatureType.VISUAL, shape=(3, IMAGE_SIZE, IMAGE_SIZE)),
+            OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(state_dim,)),
+        },
+        output_features={
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(action_dim,)),
+        },
+        device="cpu",
+        chunk_size=action_horizon,
+        n_action_steps=min(N_ACTION_STEPS, action_horizon),
+        action_dim=action_dim,
+        state_dim=state_dim,
+        num_video_frames=num_video_frames,
+        num_action_tokens_per_timestep=2,
+        num_embodied_action_tokens_per_instruction=3,
+        num_inference_timesteps=2,
+        action_hidden_size=QWEN_HIDDEN_SIZE,
+        action_model_type="DiT-test",
+        action_num_layers=1,
+        predictor_depth=1,
+        predictor_num_heads=2,
+        predictor_mlp_ratio=2.0,
+        jepa_tubelet_size=1,
+    )
+    config.validate_features()
+    return config
+
+
+def make_train_batch(
+    batch_size: int = BATCH_SIZE,
+    action_dim: int = ACTION_DIM,
+    state_dim: int = STATE_DIM,
+    action_horizon: int = ACTION_HORIZON,
+    num_video_frames: int = NUM_VIDEO_FRAMES,
+) -> dict[str, Tensor | list[str]]:
+    return {
+        f"{OBS_IMAGES}.laptop": torch.rand(batch_size, num_video_frames, 3, IMAGE_SIZE, IMAGE_SIZE),
+        OBS_STATE: torch.randn(batch_size, 1, state_dim),
+        ACTION: torch.randn(batch_size, action_horizon, action_dim),
+        "task": ["pick up the cube"] * batch_size,
+    }
+
+
+def make_inference_batch(
+    batch_size: int = BATCH_SIZE,
+    state_dim: int = STATE_DIM,
+) -> dict[str, Tensor | list[str]]:
+    return {
+        f"{OBS_IMAGES}.laptop": torch.rand(batch_size, 3, IMAGE_SIZE, IMAGE_SIZE),
+        OBS_STATE: torch.randn(batch_size, state_dim),
+        "task": ["pick up the cube"] * batch_size,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Fake external models (replace Qwen3-VL and V-JEPA at test time)
+# ---------------------------------------------------------------------------
+
+
+class _FakeLanguageLayer(nn.Module):
+    """Leaf module whose forward hook is captured by _qwen_last_decoder_hidden."""
+
+    def __init__(self, hidden_size: int) -> None:
+        super().__init__()
+        self._hidden_size = hidden_size
+
+    def forward(self, hidden: Tensor, **_: object) -> tuple[Tensor, ...]:
+        return (hidden,)
+
+
+class _FakeLanguageModel(nn.Module):
+    def __init__(self, hidden_size: int) -> None:
+        super().__init__()
+        self._hidden_size = hidden_size
+        self.layers = nn.ModuleList([_FakeLanguageLayer(hidden_size)])
+
+    def forward(self, input_ids: Tensor, **_: object) -> SimpleNamespace:
+        batch_size, seq_len = input_ids.shape
+        hidden = torch.zeros(batch_size, seq_len, self._hidden_size, device=input_ids.device)
+        self.layers[-1](hidden)
+        return SimpleNamespace()
+
+
+class _FakeQwenInnerModel(nn.Module):
+    """Mimics the `.model.model` level that _qwen_last_decoder_hidden walks into."""
+
+    def __init__(self, hidden_size: int) -> None:
+        super().__init__()
+        self.language_model = _FakeLanguageModel(hidden_size)
+
+    def forward(self, input_ids: Tensor, **kwargs: object) -> SimpleNamespace:
+        return self.language_model(input_ids)
+
+
+class _FakeQwenBackbone(nn.Module):
+    def __init__(self, hidden_size: int) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(1))
+        self.config = SimpleNamespace(
+            hidden_size=hidden_size,
+            text_config=SimpleNamespace(hidden_size=hidden_size),
+        )
+        self.model = _FakeQwenInnerModel(hidden_size)
+
+    @property
+    def device(self) -> torch.device:
+        return self.weight.device
+
+    def forward(self, input_ids: Tensor, **_: object) -> SimpleNamespace:
+        batch_size, seq_len = input_ids.shape
+        hidden_size = self.config.hidden_size
+        values = torch.arange(
+            batch_size * seq_len * hidden_size,
+            device=input_ids.device,
+            dtype=torch.float32,
+        ).view(batch_size, seq_len, hidden_size)
+        hidden = values / values.numel() + self.weight
+        self.model(input_ids)  # call through so the forward hook on layers[-1] fires
+        return SimpleNamespace(hidden_states=[hidden])
+
+
+class _FakeQwenInterface(nn.Module):
+    def __init__(self, config: VLAJEPAConfig) -> None:
+        super().__init__()
+        self.config = config
+        self.model = _FakeQwenBackbone(hidden_size=QWEN_HIDDEN_SIZE)
+
+    @staticmethod
+    def _get_torch_dtype(dtype_name: str) -> torch.dtype:
+        return torch.float32 if dtype_name == "float32" else torch.bfloat16
+
+    def expand_tokenizer(self) -> tuple[list[str], list[int], int]:
+        max_action_tokens = self.config.chunk_size * self.config.num_action_tokens_per_timestep
+        action_tokens = [self.config.special_action_token.format(idx) for idx in range(max_action_tokens)]
+        action_token_ids = list(range(1000, 1000 + max_action_tokens))
+        return action_tokens, action_token_ids, 2000
+
+    def build_inputs(
+        self,
+        images: list[list[Image.Image]],
+        instructions: list[str],
+        action_prompt: str,
+        embodied_prompt: str,
+    ) -> dict[str, Tensor]:
+        batch_size = len(images)
+        del images, instructions, action_prompt, embodied_prompt
+        action_count = (self.config.num_video_frames - 1) * self.config.num_action_tokens_per_timestep
+        token_ids = (
+            [10]
+            + list(range(1000, 1000 + action_count))
+            + [2000] * self.config.num_embodied_action_tokens_per_instruction
+            + [11]
+        )
+        return {
+            "input_ids": torch.tensor(
+                [token_ids] * batch_size,
+                device=self.model.device,
+                dtype=torch.long,
+            )
+        }
+
+    @staticmethod
+    def tensor_to_pil(image_tensor: Tensor) -> Image.Image:
+        image = image_tensor.detach().cpu()
+        if image.ndim == 3 and image.shape[0] in (1, 3):
+            image = image.permute(1, 2, 0)
+        image = (image.float().clamp(0, 1) * 255).to(torch.uint8).numpy()
+        return Image.fromarray(image)
+
+
+class _FakeVideoEncoder(nn.Module):
+    def __init__(self, hidden_size: int = 8, tubelet_size: int = 1) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(1))
+        # image_size must be >= patch_size (16) so the predictor grid is non-zero.
+        # Setting image_size=16 gives a 1x1 grid (1 patch per frame).
+        self.config = SimpleNamespace(hidden_size=hidden_size, tubelet_size=tubelet_size, image_size=16)
+
+    @property
+    def device(self) -> torch.device:
+        return self.weight.device
+
+    def get_vision_features(self, pixel_values_videos: Tensor) -> Tensor:
+        batch_size, num_frames = pixel_values_videos.shape[:2]
+        hidden_size = self.config.hidden_size
+        frame_values = pixel_values_videos.float().mean(dim=(2, 3, 4), keepdim=False)
+        return frame_values[:, :, None].expand(batch_size, num_frames, hidden_size)
+
+
+class _FakeVideoProcessor:
+    def __call__(self, videos, return_tensors: str) -> dict[str, Tensor]:
+        assert return_tensors == "pt"
+        if isinstance(videos, list):
+            pixel_values = torch.stack([torch.as_tensor(v) for v in videos])
+        else:
+            pixel_values = torch.as_tensor(videos).unsqueeze(0)
+        return {"pixel_values_videos": pixel_values}
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def patch_vla_jepa_external_models(monkeypatch: pytest.MonkeyPatch) -> None:
+    from lerobot.policies.vla_jepa import modeling_vla_jepa
+
+    monkeypatch.setattr(modeling_vla_jepa, "Qwen3VLInterface", _FakeQwenInterface)
+    monkeypatch.setattr(
+        modeling_vla_jepa.AutoModel,
+        "from_pretrained",
+        lambda *args, **kwargs: _FakeVideoEncoder(),
+    )
+    monkeypatch.setattr(
+        modeling_vla_jepa.AutoVideoProcessor,
+        "from_pretrained",
+        lambda *args, **kwargs: _FakeVideoProcessor(),
+    )

tests/policies/vla_jepa/test_action_head.py

@@ -0,0 +1,157 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import pytest
+import torch
+
+pytest.importorskip("diffusers")
+
+from conftest import (
+    ACTION_DIM,
+    ACTION_HORIZON,
+    BATCH_SIZE,
+    QWEN_HIDDEN_SIZE,
+    STATE_DIM,
+    make_config,
+    set_seed_all,
+)  # noqa: E402
+
+from lerobot.policies.vla_jepa.action_head import (  # noqa: E402
+    VLAJEPAActionHead,
+)
+
+# ---------------------------------------------------------------------------
+# VLAJEPAActionHead
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),  # default test dims
+        (7, 0, 16),  # no proprioceptive state, production-like action space
+        (6, 8, 8),  # medium dims
+    ],
+)
+def test_action_head_sample_time_range(action_dim: int, state_dim: int, action_horizon: int) -> None:
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    t = head.sample_time(batch_size=200, device=torch.device("cpu"), dtype=torch.float32)
+    assert t.shape == (200,)
+    assert torch.isfinite(t).all()
+
+
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),
+        (7, 0, 16),
+        (6, 8, 8),
+    ],
+)
+def test_action_head_build_inputs_shape(action_dim: int, state_dim: int, action_horizon: int) -> None:
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(2, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(2, action_horizon, action_dim)
+    timesteps = torch.randint(0, 100, (2,))
+
+    state = torch.randn(2, state_dim) if state_dim > 0 else None
+    out_with = head._build_inputs(conditioning, actions, state, timesteps)
+    out_none = head._build_inputs(conditioning, actions, None, timesteps)
+
+    assert out_with.ndim == 3 and out_none.ndim == 3
+    if state_dim > 0:
+        assert out_with.shape[1] > out_none.shape[1]
+    assert torch.isfinite(out_with).all() and torch.isfinite(out_none).all()
+
+
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),
+        (7, 0, 16),
+        (6, 8, 8),
+    ],
+)
+def test_action_head_forward_loss_valid(action_dim: int, state_dim: int, action_horizon: int) -> None:
+    set_seed_all(42)
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(2, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(2, action_horizon, action_dim)
+    state = torch.randn(2, state_dim) if state_dim > 0 else None
+    loss = head.forward(conditioning, actions, state)
+    assert loss.shape == ()
+    assert torch.isfinite(loss) and loss > 0
+
+
+def test_action_head_forward_gradient_flows() -> None:
+    set_seed_all(42)
+    config = make_config()
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(BATCH_SIZE, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(BATCH_SIZE, ACTION_HORIZON, ACTION_DIM)
+    state = torch.randn(BATCH_SIZE, STATE_DIM)
+    loss = head.forward(conditioning, actions, state)
+    loss.backward()
+    assert any(p.grad is not None for p in head.parameters() if p.requires_grad)
+
+
+@torch.no_grad()
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),
+        (7, 0, 16),
+        (6, 8, 8),
+    ],
+)
+def test_action_head_predict_action_shape(action_dim: int, state_dim: int, action_horizon: int) -> None:
+    set_seed_all(42)
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(2, 4, QWEN_HIDDEN_SIZE)
+    state = torch.randn(2, state_dim) if state_dim > 0 else None
+    pred = head.predict_action(conditioning, state)
+    assert tuple(pred.shape) == (2, action_horizon, action_dim)
+    assert torch.isfinite(pred).all()
+
+
+# ---------------------------------------------------------------------------
+# action_is_pad masking
+# ---------------------------------------------------------------------------
+
+
+def test_action_head_loss_fully_padded_is_zero() -> None:
+    """Loss is 0 when every timestep is padded (exercises the clamp_min guard)."""
+    set_seed_all(42)
+    config = make_config()
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(BATCH_SIZE, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(BATCH_SIZE, ACTION_HORIZON, ACTION_DIM)
+    state = torch.randn(BATCH_SIZE, STATE_DIM)
+
+    action_is_pad = torch.ones(BATCH_SIZE, ACTION_HORIZON, dtype=torch.bool)
+    loss = head.forward(conditioning, actions, state, action_is_pad)
+    assert loss.item() == 0.0
+
+
+def test_action_head_loss_none_matches_no_padding() -> None:
+    """action_is_pad=None is equivalent to an all-False (no padding) mask."""
+    set_seed_all(42)
+    config = make_config()
+    head = VLAJEPAActionHead(config, cross_attention_dim=QWEN_HIDDEN_SIZE)
+    conditioning = torch.randn(BATCH_SIZE, 4, QWEN_HIDDEN_SIZE)
+    actions = torch.randn(BATCH_SIZE, ACTION_HORIZON, ACTION_DIM)
+    state = torch.randn(BATCH_SIZE, STATE_DIM)
+
+    set_seed_all(0)
+    loss_none = head.forward(conditioning, actions, state, action_is_pad=None)
+
+    set_seed_all(0)
+    no_pad = torch.zeros(BATCH_SIZE, ACTION_HORIZON, dtype=torch.bool)
+    loss_zeros = head.forward(conditioning, actions, state, action_is_pad=no_pad)
+
+    assert torch.isclose(loss_none, loss_zeros)

tests/policies/vla_jepa/test_configuration.py

@@ -0,0 +1,57 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import pytest
+from conftest import ACTION_DIM, ACTION_HORIZON, IMAGE_SIZE, NUM_VIDEO_FRAMES, STATE_DIM, make_config
+
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig
+from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE
+
+
+def test_delta_indices() -> None:
+    config = make_config()
+    assert config.observation_delta_indices == list(range(NUM_VIDEO_FRAMES))
+    assert config.action_delta_indices == list(range(ACTION_HORIZON))
+
+
+def test_n_action_steps_exceeds_chunk_size_raises() -> None:
+    with pytest.raises(ValueError, match="n_action_steps"):
+        VLAJEPAConfig(chunk_size=4, n_action_steps=8)
+
+
+def test_too_few_video_frames_raises() -> None:
+    with pytest.raises(ValueError, match="video_horizon"):
+        VLAJEPAConfig(
+            chunk_size=16,
+            n_action_steps=16,
+            num_video_frames=2,
+            jepa_tubelet_size=2,  # needs >= 4 frames (2 for current, 2 for future) to have a window of size > 0
+        )
+
+
+def test_validate_features_no_image_raises() -> None:
+    config = VLAJEPAConfig(
+        input_features={OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(STATE_DIM,))},
+        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,))},
+    )
+    with pytest.raises(ValueError, match="at least one visual input feature"):
+        config.validate_features()
+
+
+def test_validate_features_no_action_raises() -> None:
+    config = VLAJEPAConfig(
+        input_features={
+            f"{OBS_IMAGES}.cam": PolicyFeature(type=FeatureType.VISUAL, shape=(3, IMAGE_SIZE, IMAGE_SIZE)),
+        },
+        output_features={},
+    )
+    with pytest.raises(ValueError, match="action output feature"):
+        config.validate_features()
+
+
+def test_validate_features_sets_action_dim_from_feature() -> None:
+    config = make_config(action_dim=6, state_dim=10)
+    assert config.action_dim == 6
+    assert config.state_dim == 10

tests/policies/vla_jepa/test_vla_jepa.py

@@ -0,0 +1,598 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import os
+from copy import deepcopy
+
+import numpy as np
+import pytest
+import torch
+from torch import Tensor
+
+pytest.importorskip("transformers")
+pytest.importorskip("diffusers")
+
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:In CPU autocast, but the target dtype is not supported:UserWarning"
+)
+
+from conftest import (  # noqa: E402
+    ACTION_DIM,
+    ACTION_HORIZON,
+    BATCH_SIZE,
+    EXPECTED_ACTION_CHUNK_SHAPE,
+    EXPECTED_SELECT_ACTION_SHAPE,
+    IMAGE_SIZE,
+    N_ACTION_STEPS,
+    QWEN_HIDDEN_SIZE,
+    STATE_DIM,
+    make_config,
+    make_inference_batch,
+    make_train_batch,
+    set_seed_all,
+)
+
+from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig  # noqa: E402
+from lerobot.policies.vla_jepa.modeling_vla_jepa import VLAJEPAPolicy  # noqa: E402
+from lerobot.utils.constants import ACTION  # noqa: E402
+
+PRETRAINED_REPO_ID = "ginwind/VLA-JEPA"
+PRETRAINED_SUBFOLDER = "LIBERO"
+
+# extended hub tests load the full converted safetensors checkpoints (~5 GB) and are
+# skipped by default.  Set VLA_JEPA_EXTENDED=1 to opt in.
+_VLA_JEPA_EXTENDED = os.environ.get("VLA_JEPA_EXTENDED", "0") != "0"
+extended_test = pytest.mark.skipif(not _VLA_JEPA_EXTENDED, reason="Set VLA_JEPA_EXTENDED=1 to run hub tests")
+
+
+# ---------------------------------------------------------------------------
+# Core training / inference tests
+# ---------------------------------------------------------------------------
+
+
+def test_training_forward_pass(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.train()
+
+    batch = make_train_batch()
+    batch_before = deepcopy(batch)
+
+    loss, logs = policy.forward(batch)
+
+    assert loss.shape == ()
+    assert torch.isfinite(loss)
+    assert set(logs) == {"action_loss", "wm_loss", "loss"}
+    assert logs["action_loss"] > 0
+    assert logs["wm_loss"] >= 0
+
+    loss.backward()
+    assert any(p.grad is not None for p in policy.model.action_model.parameters() if p.requires_grad)
+    # Batch must not be mutated.
+    assert set(batch) == set(batch_before)
+    for key, value in batch.items():
+        if isinstance(value, Tensor):
+            assert torch.equal(value, batch_before[key])
+        else:
+            assert value == batch_before[key]
+
+
+@pytest.mark.parametrize("batch_size", [1, 2, 4])
+def test_training_forward_various_batch_sizes(patch_vla_jepa_external_models: None, batch_size: int) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.train()
+    loss, logs = policy.forward(make_train_batch(batch_size=batch_size))
+    assert torch.isfinite(loss) and loss > 0
+    assert set(logs) == {"action_loss", "wm_loss", "loss"}
+
+
+@pytest.mark.parametrize(
+    "action_dim,state_dim,action_horizon",
+    [
+        (3, 4, 4),
+        (7, 0, 16),
+        (6, 8, 8),
+    ],
+)
+def test_training_forward_various_dims(
+    patch_vla_jepa_external_models: None,
+    action_dim: int,
+    state_dim: int,
+    action_horizon: int,
+) -> None:
+    set_seed_all(42)
+    config = make_config(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    policy = VLAJEPAPolicy(config)
+    policy.train()
+    batch = make_train_batch(action_dim=action_dim, state_dim=state_dim, action_horizon=action_horizon)
+    loss, _ = policy.forward(batch)
+    assert torch.isfinite(loss) and loss > 0
+
+
+@torch.no_grad()
+def test_action_generation_shape(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    batch = make_inference_batch()
+
+    chunk = policy.predict_action_chunk(batch)
+    assert tuple(chunk.shape) == EXPECTED_ACTION_CHUNK_SHAPE
+    assert chunk.device.type == "cpu"
+    assert torch.isfinite(chunk).all()
+
+    a1 = policy.select_action(batch)
+    a2 = policy.select_action(batch)
+    assert tuple(a1.shape) == EXPECTED_SELECT_ACTION_SHAPE
+    assert tuple(a2.shape) == EXPECTED_SELECT_ACTION_SHAPE
+    assert torch.isfinite(a1).all() and torch.isfinite(a2).all()
+
+
+@torch.no_grad()
+@pytest.mark.parametrize("action_dim,state_dim", [(3, 4), (7, 0), (6, 8)])
+def test_action_generation_various_dims(
+    patch_vla_jepa_external_models: None, action_dim: int, state_dim: int
+) -> None:
+    set_seed_all(42)
+    config = make_config(action_dim=action_dim, state_dim=state_dim)
+    policy = VLAJEPAPolicy(config)
+    policy.eval()
+    batch = make_inference_batch(state_dim=state_dim)
+    chunk = policy.predict_action_chunk(batch)
+    assert chunk.shape[-1] == action_dim
+    assert torch.isfinite(chunk).all()
+
+
+@torch.no_grad()
+def test_inference_reproducibility(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    batch = make_inference_batch()
+
+    set_seed_all(123)
+    actions_1 = policy.predict_action_chunk(batch)
+    set_seed_all(123)
+    actions_2 = policy.predict_action_chunk(batch)
+
+    assert tuple(actions_1.shape) == EXPECTED_ACTION_CHUNK_SHAPE
+    assert torch.allclose(actions_1, actions_2, atol=1e-6)
+
+
+@torch.no_grad()
+def test_predict_action_chunk_always_finite(patch_vla_jepa_external_models: None) -> None:
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    for seed in [0, 42, 123]:
+        set_seed_all(seed)
+        chunk = policy.predict_action_chunk(make_inference_batch())
+        assert torch.isfinite(chunk).all(), f"non-finite actions with seed={seed}"
+
+
+# ---------------------------------------------------------------------------
+# Action queue behaviour
+# ---------------------------------------------------------------------------
+
+
+@torch.no_grad()
+def test_select_action_queue_drains_before_refill(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    batch = make_inference_batch()
+
+    # First call fills the queue (n_action_steps items) and pops one.
+    a1 = policy.select_action(batch)
+    assert len(policy._queues[ACTION]) == N_ACTION_STEPS - 1
+
+    # Second call pops from the existing queue without calling predict_action_chunk.
+    a2 = policy.select_action(batch)
+    assert tuple(a1.shape) == EXPECTED_SELECT_ACTION_SHAPE
+    assert tuple(a2.shape) == EXPECTED_SELECT_ACTION_SHAPE
+
+
+@torch.no_grad()
+def test_reset_clears_action_queue(patch_vla_jepa_external_models: None) -> None:
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(make_config())
+    policy.eval()
+    policy.select_action(make_inference_batch())
+    assert len(policy._queues[ACTION]) > 0
+
+    policy.reset()
+    assert len(policy._queues[ACTION]) == 0
+
+
+# ---------------------------------------------------------------------------
+# Format conversion
+# ---------------------------------------------------------------------------
+
+
+def test_prepare_model_inputs_training_format(patch_vla_jepa_external_models: None) -> None:
+    from PIL import Image
+
+    policy = VLAJEPAPolicy(make_config())
+    examples = policy._prepare_model_inputs(make_train_batch())
+
+    assert len(examples) == BATCH_SIZE
+    for ex in examples:
+        assert set(ex) >= {"image", "video", "lang", "action", "state"}
+        assert len(ex["image"]) == 1 and isinstance(ex["image"][0], Image.Image)
+        assert ex["video"].ndim == 5 and ex["video"].dtype == np.uint8  # [V,T,H,W,C]
+        assert ex["action"].shape == (ACTION_HORIZON, ACTION_DIM)
+        assert ex["state"].shape == (1, STATE_DIM)
+
+
+def test_prepare_model_inputs_inference_omits_action(patch_vla_jepa_external_models: None) -> None:
+    policy = VLAJEPAPolicy(make_config())
+    for ex in policy._prepare_model_inputs(make_inference_batch()):
+        assert "action" not in ex
+        assert "image" in ex and "video" in ex and "lang" in ex
+
+
+def test_prepare_model_inputs_missing_task_uses_default(patch_vla_jepa_external_models: None) -> None:
+    policy = VLAJEPAPolicy(make_config())
+    batch = make_inference_batch()
+    del batch["task"]
+    examples = policy._prepare_model_inputs(batch)
+    assert all(isinstance(ex["lang"], str) and len(ex["lang"]) > 0 for ex in examples)
+
+
+def test_prepare_model_inputs_string_task_broadcast(patch_vla_jepa_external_models: None) -> None:
+    policy = VLAJEPAPolicy(make_config())
+    batch = make_inference_batch()
+    batch["task"] = "open the drawer"
+    assert all(ex["lang"] == "open the drawer" for ex in policy._prepare_model_inputs(batch))
+
+
+def test_prepare_model_inputs_no_state_omitted(patch_vla_jepa_external_models: None) -> None:
+    from lerobot.utils.constants import OBS_STATE
+
+    policy = VLAJEPAPolicy(make_config())
+    batch = make_inference_batch()
+    del batch[OBS_STATE]
+    assert all("state" not in ex for ex in policy._prepare_model_inputs(batch))
+
+
+# ---------------------------------------------------------------------------
+# Pretrained checkpoint
+# Hub tests (opt-in: VLA_JEPA_EXTENDED=1)
+# ---------------------------------------------------------------------------
+
+
+def _make_hub_train_batch(policy: VLAJEPAPolicy, batch_size: int = 1) -> dict:
+    """Build a training batch whose keys/shapes match a hub-loaded policy config."""
+    cfg = policy.config
+    batch: dict = {"task": ["pick up the cube"] * batch_size}
+    for key, feat in cfg.image_features.items():
+        h, w = feat.shape[-2], feat.shape[-1]
+        batch[key] = torch.rand(batch_size, cfg.num_video_frames, 3, h, w)
+    if cfg.robot_state_feature is not None:
+        batch["observation.state"] = torch.randn(batch_size, 1, cfg.robot_state_feature.shape[0])
+    batch[ACTION] = torch.randn(batch_size, cfg.chunk_size, cfg.action_dim)
+    return batch
+
+
+def _make_hub_inference_batch(policy: VLAJEPAPolicy, batch_size: int = 1) -> dict:
+    """Build an inference batch whose keys/shapes match a hub-loaded policy config."""
+    cfg = policy.config
+    batch: dict = {"task": ["pick up the cube"] * batch_size}
+    for key, feat in cfg.image_features.items():
+        h, w = feat.shape[-2], feat.shape[-1]
+        batch[key] = torch.rand(batch_size, 3, h, w)
+    if cfg.robot_state_feature is not None:
+        batch["observation.state"] = torch.randn(batch_size, cfg.robot_state_feature.shape[0])
+    return batch
+
+
+_CP_ROOT = "lerobot"
+
+# Each tuple: (repo_id, enable_world_model)
+_HUB_VARIANTS = [
+    (f"{_CP_ROOT}/VLA-JEPA-LIBERO", True),
+    (f"{_CP_ROOT}/VLA-JEPA-Pretrain", True),
+    (f"{_CP_ROOT}/VLA-JEPA-SimplerEnv", False),
+]
+
+
+@extended_test
+@pytest.mark.parametrize("repo_id,enable_world_model", _HUB_VARIANTS)
+def test_hub_checkpoint_loads(repo_id: str, enable_world_model: bool) -> None:
+    """Policy loads from the converted safetensors checkpoint on the Hub."""
+    policy = VLAJEPAPolicy.from_pretrained(repo_id)
+    assert policy.config.enable_world_model == enable_world_model
+    assert sum(p.numel() for p in policy.parameters()) > 0
+
+
+@extended_test
+@pytest.mark.parametrize("repo_id,enable_world_model", _HUB_VARIANTS)
+def test_hub_checkpoint_forward_pass(repo_id: str, enable_world_model: bool) -> None:
+    """Policy loaded from hub produces finite losses with a correctly-shaped batch."""
+    policy = VLAJEPAPolicy.from_pretrained(repo_id)
+    policy.train()
+
+    batch = _make_hub_train_batch(policy)
+    loss, logs = policy.forward(batch)
+    assert torch.isfinite(loss)
+    assert "action_loss" in logs
+    if enable_world_model:
+        assert "wm_loss" in logs
+
+
+@extended_test
+def test_hub_freeze_qwen_disables_world_model() -> None:
+    """freeze_qwen=True (via cli_overrides) freezes qwen and disables the world model."""
+    policy = VLAJEPAPolicy.from_pretrained(f"{_CP_ROOT}/VLA-JEPA-LIBERO", cli_overrides=["freeze_qwen=true"])
+    assert not policy.config.enable_world_model
+    assert policy.model.video_predictor is None
+    qwen_params = list(policy.model.qwen.parameters())
+    assert all(not p.requires_grad for p in qwen_params)
+    assert any(p.requires_grad for p in policy.model.action_model.parameters())
+
+
+@extended_test
+def test_hub_disable_world_model_loads_simpler_env() -> None:
+    """SimplerEnv checkpoint (world model disabled) loads cleanly and runs inference."""
+    policy = VLAJEPAPolicy.from_pretrained(f"{_CP_ROOT}/VLA-JEPA-SimplerEnv")
+    assert not policy.config.enable_world_model
+    assert policy.model.video_predictor is None
+    assert policy.model.video_encoder is None
+
+
+@extended_test
+def test_hub_libero_inference_shape() -> None:
+    """select_action returns the expected shape using the LIBERO hub checkpoint."""
+    policy = VLAJEPAPolicy.from_pretrained(f"{_CP_ROOT}/VLA-JEPA-LIBERO")
+    policy.eval()
+    batch = _make_hub_inference_batch(policy)
+    action = policy.select_action(batch)
+    assert action.shape[-1] == policy.config.action_dim
+
+
+# ---------------------------------------------------------------------------
+# Postprocessor unnormalization tests
+#
+# These tests verify that the postprocessor pipeline (clip → unnorm → binarize)
+# correctly applies MIN_MAX unnormalization after predict_action_chunk.
+# ---------------------------------------------------------------------------
+
+
+def _make_dataset_stats(action_dim: int = ACTION_DIM) -> dict:
+    """Returns sample dataset_stats with a simple [i, i+10] range per action dim."""
+    from lerobot.utils.constants import ACTION
+
+    return {
+        ACTION: {
+            "min": torch.tensor([float(i) for i in range(action_dim)], dtype=torch.float32),
+            "max": torch.tensor([float(i) + 10.0 for i in range(action_dim)], dtype=torch.float32),
+        }
+    }
+
+
+@torch.no_grad()
+def test_postprocessor_unnormalizes_actions(patch_vla_jepa_external_models: None) -> None:
+    """UnnormalizerProcessorStep with MIN_MAX produces the correct inverse of MIN_MAX normalization."""
+    from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+    from lerobot.processor import UnnormalizerProcessorStep
+    from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+    from lerobot.utils.constants import ACTION
+
+    dataset_stats = _make_dataset_stats()
+
+    rng = np.random.default_rng(7)
+    actions_np = rng.uniform(-1.0, 1.0, (2, ACTION_HORIZON, ACTION_DIM)).astype(np.float32)
+
+    a_min = dataset_stats[ACTION]["min"].numpy()
+    a_max = dataset_stats[ACTION]["max"].numpy()
+    expected = (actions_np + 1.0) / 2.0 * (a_max - a_min) + a_min
+
+    features = {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,))}
+    unnorm_step = UnnormalizerProcessorStep(
+        features=features,
+        norm_map={FeatureType.ACTION: NormalizationMode.MIN_MAX},
+        stats=dataset_stats,
+    )
+
+    actions_tensor = torch.from_numpy(actions_np)
+    transition = policy_action_to_transition(actions_tensor)
+    result = transition_to_policy_action(unnorm_step(transition)).numpy()
+
+    np.testing.assert_allclose(result, expected, rtol=1e-5, atol=1e-6)
+
+
+@torch.no_grad()
+def test_postprocessor_clip_clamps_before_unnorm(patch_vla_jepa_external_models: None) -> None:
+    """ClipActionsProcessorStep clamps to [-1, 1] before unnormalization."""
+    from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+    from lerobot.policies.vla_jepa.processor_vla_jepa import ClipActionsProcessorStep
+    from lerobot.processor import UnnormalizerProcessorStep
+    from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+    from lerobot.utils.constants import ACTION
+
+    dataset_stats = _make_dataset_stats()
+    a_min = dataset_stats[ACTION]["min"].numpy()
+    a_max = dataset_stats[ACTION]["max"].numpy()
+
+    # Deliberately out-of-range inputs
+    actions_np = np.array([[[2.0] * ACTION_DIM, [-3.0] * ACTION_DIM]], dtype=np.float32)
+    clipped = np.clip(actions_np, -1.0, 1.0)
+    expected = (clipped + 1.0) / 2.0 * (a_max - a_min) + a_min
+
+    features = {ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,))}
+    clip_step = ClipActionsProcessorStep()
+    unnorm_step = UnnormalizerProcessorStep(
+        features=features,
+        norm_map={FeatureType.ACTION: NormalizationMode.MIN_MAX},
+        stats=dataset_stats,
+    )
+
+    transition = policy_action_to_transition(torch.from_numpy(actions_np))
+    transition = clip_step(transition)
+    result = transition_to_policy_action(unnorm_step(transition)).numpy()
+
+    np.testing.assert_allclose(result, expected, rtol=1e-5, atol=1e-6)
+
+
+@torch.no_grad()
+def test_postprocessor_applied_after_predict_action_chunk(
+    patch_vla_jepa_external_models: None, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """predict_action_chunk returns raw actions; the postprocessor applies unnormalization.
+
+    Verifies the split: predict_action_chunk returns normalized actions, and calling the
+    postprocessor on them produces the correctly unnormalized result.
+    """
+    from lerobot.policies.vla_jepa.processor_vla_jepa import make_vla_jepa_pre_post_processors
+
+    raw_actions = np.zeros((BATCH_SIZE, ACTION_HORIZON, ACTION_DIM), dtype=np.float32)
+
+    cfg = make_config()
+    cfg.clip_normalized_actions = False
+    cfg.binarize_gripper_action = False
+    policy = VLAJEPAPolicy(cfg)
+    policy.eval()
+    monkeypatch.setattr(policy.model, "predict_action", lambda *a, **kw: raw_actions.copy())
+
+    dataset_stats = _make_dataset_stats()
+    _, postprocessor = make_vla_jepa_pre_post_processors(cfg, dataset_stats)
+
+    batch = make_inference_batch()
+    chunk = policy.predict_action_chunk(batch)
+
+    # predict_action_chunk returns raw (normalized) actions
+    assert torch.allclose(chunk, torch.zeros_like(chunk), atol=1e-6), (
+        "predict_action_chunk should return raw actions without unnormalization applied."
+    )
+
+    # Postprocessor applies unnormalization: 0 → (0+1)/2 * (max-min) + min = 5 + i
+    unnormed = postprocessor(chunk)
+    from lerobot.utils.constants import ACTION
+
+    a_min = dataset_stats[ACTION]["min"].numpy()
+    a_max = dataset_stats[ACTION]["max"].numpy()
+    expected_first = 0.5 * (0.0 + 1.0) * (a_max[0] - a_min[0]) + a_min[0]
+    assert unnormed[0, 0, 0].item() == pytest.approx(expected_first, abs=1e-5)
+
+
+# ---------------------------------------------------------------------------
+# World-model view adjustment (padding / trimming) tests
+# ---------------------------------------------------------------------------
+
+
+_MULTIVIEW_NUM_FRAMES = 4  # must be >= 2 * jepa_tubelet_size (=2) for world-model tests
+
+
+def _make_multiview_config(num_views: int, jepa_tubelet_size: int = 2) -> VLAJEPAConfig:
+    from lerobot.configs.types import FeatureType, PolicyFeature
+    from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
+
+    config = VLAJEPAConfig(
+        input_features={
+            **{
+                f"{OBS_IMAGES}.cam{i}": PolicyFeature(
+                    type=FeatureType.VISUAL, shape=(3, IMAGE_SIZE, IMAGE_SIZE)
+                )
+                for i in range(num_views)
+            },
+            OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(STATE_DIM,)),
+        },
+        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(ACTION_DIM,))},
+        device="cpu",
+        chunk_size=ACTION_HORIZON,
+        n_action_steps=N_ACTION_STEPS,
+        action_dim=ACTION_DIM,
+        state_dim=STATE_DIM,
+        num_video_frames=_MULTIVIEW_NUM_FRAMES,
+        num_action_tokens_per_timestep=2,
+        num_embodied_action_tokens_per_instruction=3,
+        num_inference_timesteps=2,
+        action_hidden_size=QWEN_HIDDEN_SIZE,
+        action_model_type="DiT-test",
+        action_num_layers=1,
+        predictor_depth=1,
+        predictor_num_heads=2,
+        predictor_mlp_ratio=2.0,
+        jepa_tubelet_size=jepa_tubelet_size,
+    )
+    config.validate_features()
+    return config
+
+
+def _make_multiview_train_batch(num_views: int, batch_size: int = BATCH_SIZE) -> dict:
+    from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
+
+    batch = {
+        f"{OBS_IMAGES}.cam{i}": torch.rand(batch_size, _MULTIVIEW_NUM_FRAMES, 3, IMAGE_SIZE, IMAGE_SIZE)
+        for i in range(num_views)
+    }
+    batch[OBS_STATE] = torch.randn(batch_size, 1, STATE_DIM)
+    batch[ACTION] = torch.randn(batch_size, ACTION_HORIZON, ACTION_DIM)
+    batch["task"] = ["pick up the cube"] * batch_size
+    return batch
+
+
+@pytest.mark.parametrize(
+    "num_views",
+    [
+        1,  # fewer views than jepa_tubelet_size → first view duplicated
+        2,  # exact match → unchanged
+        3,  # more views than jepa_tubelet_size → trimmed to first two
+    ],
+)
+def test_training_forward_world_model_view_adjustment(
+    patch_vla_jepa_external_models: None,
+    num_views: int,
+) -> None:
+    """World-model view padding/trimming must not break the training forward pass."""
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(_make_multiview_config(num_views=num_views, jepa_tubelet_size=2))
+    policy.train()
+    loss, logs = policy.forward(_make_multiview_train_batch(num_views=num_views))
+    assert torch.isfinite(loss)
+    assert logs["wm_loss"] >= 0
+
+
+def test_single_view_is_duplicated_for_world_model(patch_vla_jepa_external_models: None) -> None:
+    """With one dataset view and jepa_tubelet_size=2, the view must be duplicated before encoding."""
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(_make_multiview_config(num_views=1, jepa_tubelet_size=2))
+    policy.train()
+
+    captured_videos: list = []
+    original_processor = policy.model.video_processor
+
+    class _CapturingProcessor:
+        def __call__(self, videos: list, return_tensors: str) -> dict:
+            captured_videos.extend(videos)
+            return original_processor(videos=videos, return_tensors=return_tensors)
+
+    policy.model.video_processor = _CapturingProcessor()
+    policy.forward(_make_multiview_train_batch(num_views=1))
+
+    # reshape is batch-major: (b0v0, b0v1, b1v0, b1v1, …)
+    assert len(captured_videos) == BATCH_SIZE * 2
+    for i in range(BATCH_SIZE):
+        np.testing.assert_array_equal(captured_videos[2 * i], captured_videos[2 * i + 1])
+
+
+def test_excess_views_trimmed_for_world_model(patch_vla_jepa_external_models: None) -> None:
+    """With three dataset views and jepa_tubelet_size=2, only the first two views reach the encoder."""
+    set_seed_all(42)
+    policy = VLAJEPAPolicy(_make_multiview_config(num_views=3, jepa_tubelet_size=2))
+    policy.train()
+
+    captured_videos: list = []
+    original_processor = policy.model.video_processor
+
+    class _CapturingProcessor:
+        def __call__(self, videos: list, return_tensors: str) -> dict:
+            captured_videos.extend(videos)
+            return original_processor(videos=videos, return_tensors=return_tensors)
+
+    policy.model.video_processor = _CapturingProcessor()
+    policy.forward(_make_multiview_train_batch(num_views=3))
+
+    # Only B*2 items must reach the encoder, not B*3.
+    assert len(captured_videos) == BATCH_SIZE * 2

tests/policies/vla_jepa/test_world_model.py

@@ -0,0 +1,60 @@
+#!/usr/bin/env python
+
+from __future__ import annotations
+
+import pytest
+import torch
+
+from lerobot.policies.vla_jepa.world_model import (
+    ActionConditionedVideoPredictor,
+)
+
+_ACTION_EMBED_DIM = 8
+
+
+def _make_predictor(
+    embed_dim: int = 8,
+    action_embed_dim: int = _ACTION_EMBED_DIM,
+    predictor_embed_dim: int = 24,
+    num_action_tokens: int = 2,
+    tokens_per_frame: int = 1,
+) -> ActionConditionedVideoPredictor:
+    return ActionConditionedVideoPredictor(
+        num_frames=1,
+        img_size=(1, tokens_per_frame),
+        patch_size=1,
+        tubelet_size=1,
+        embed_dim=embed_dim,
+        action_embed_dim=action_embed_dim,
+        predictor_embed_dim=predictor_embed_dim,
+        depth=1,
+        num_heads=2,
+        mlp_ratio=2.0,
+        num_action_tokens_per_step=num_action_tokens,
+    )
+
+
+@pytest.mark.parametrize(
+    "batch,num_steps,tokens_per_frame,embed_dim",
+    [
+        (1, 2, 1, 8),
+        (2, 3, 4, 8),
+        (4, 5, 2, 16),
+    ],
+)
+def test_predictor_output_shape(batch: int, num_steps: int, tokens_per_frame: int, embed_dim: int) -> None:
+    predictor = _make_predictor(
+        embed_dim=embed_dim, action_embed_dim=_ACTION_EMBED_DIM, tokens_per_frame=tokens_per_frame
+    )
+    frame_tokens = torch.randn(batch, num_steps * tokens_per_frame, embed_dim)
+    action_tokens = torch.randn(batch, num_steps * 2, _ACTION_EMBED_DIM)
+    out = predictor(frame_tokens, action_tokens)
+    assert tuple(out.shape) == (batch, num_steps * tokens_per_frame, embed_dim)
+    assert torch.isfinite(out).all()
+
+
+def test_predictor_step_mismatch_raises() -> None:
+    predictor = _make_predictor(tokens_per_frame=4)
+    frame_tokens = torch.randn(2, 3 * 4, 8)  # 3 steps, 4 tokens each
+    with pytest.raises(RuntimeError):
+        predictor(frame_tokens, torch.randn(2, 2 * 2, 8))  # 2 steps → mismatch