feat(lingbot_va): RoboTwin eef-pose eval, single-file model, Hub checkpoints

Make the LingBot-VA port runnable on both LIBERO and RoboTwin and clean up the
package to LeRobot conventions.

- Consolidate all vendored Wan2.2 model code (transformer, attention, VAE helpers,
  flow-matching scheduler, grid utils, flex-attention) into a single
  modeling_lingbot_va.py; remove the separate wan_*/schedulers modules.
- Move the fixed action (un)normalization quantiles out of the config and into the
  post-processor (LIBERO 7-DoF + RoboTwin 16-d eef); remove the conversion script in
  favour of ready-to-use LeRobot-format checkpoints on the Hub.
- Fixes found via on-sim validation: undo LIBERO's 180-degree image flip
  (image_hflip), encode obs as a multi-frame streaming-VAE clip, reset the streaming
  VAE cache between episodes, run the transformer in config.dtype, lazy-load frozen
  VAE/UMT5 by subfolder with the text encoder on CPU.
- RoboTwin: add an end-effector-pose action mode to RoboTwinEnv (16-d per-arm
  xyz+quat+gripper deltas composed onto the initial eef pose, executed via CuRobo IK)
  and the robotwin_tshape latent layout (full-res head + half-res wrists via a second
  streaming VAE) with the upstream RoboTwin action quantiles + camera mapping.
- Predicted-video saving works for both benchmarks; docs + tests updated.

Co-authored-by: Cursor <cursoragent@cursor.com>

tests/policies/lingbot_va/test_configuration.py

@@ -76,8 +76,3 @@ def test_validate_features_no_visual_raises() -> None:
 def test_invalid_attn_mode_raises() -> None:
     with pytest.raises(ValueError, match="attn_mode"):
         make_config(attn_mode="banana")
-
-
-def test_quantile_length_mismatch_raises() -> None:
-    with pytest.raises(ValueError, match="action_q01"):
-        make_config(used_action_channel_ids=[0, 1, 2], action_q01=[0.0, 0.0], action_q99=[1.0, 1.0, 1.0])

tests/policies/lingbot_va/test_factory.py

@@ -36,17 +36,3 @@ def test_get_policy_class_resolves_lazily() -> None:
     cls = get_policy_class("lingbot_va")
     assert cls.name == "lingbot_va"
     assert cls.config_class is LingBotVAConfig
-
-
-def test_convert_build_config_libero() -> None:
-    pytest.importorskip("diffusers")
-    from lerobot.policies.lingbot_va.convert_lingbot_va_checkpoints import build_config
-
-    cfg = build_config("libero", wan_pretrained_path="dummy/path", dtype="float32")
-    assert cfg.height == 128 and cfg.width == 128
-    assert cfg.used_action_channel_ids == list(range(7))
-    # validate_features (called inside build_config) must have populated the action feature.
-    from lerobot.utils.constants import ACTION
-
-    assert cfg.output_features[ACTION].shape == (7,)
-    assert len(cfg.obs_cam_keys) == 2

tests/policies/lingbot_va/test_modules.py

@@ -14,14 +14,20 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Pure-torch unit tests for the vendored LingBot-VA helper modules (no diffusers needed)."""
+"""Unit tests for the vendored LingBot-VA helper code (scheduler + grid utilities)."""
 
 from __future__ import annotations
 
+import pytest
 import torch
 
-from lerobot.policies.lingbot_va.schedulers import FlowMatchScheduler
-from lerobot.policies.lingbot_va.wan_utils import data_seq_to_patch, get_mesh_id
+pytest.importorskip("diffusers")  # the model code lives in modeling_lingbot_va, which imports diffusers
+
+from lerobot.policies.lingbot_va.modeling_lingbot_va import (  # noqa: E402
+    FlowMatchScheduler,
+    data_seq_to_patch,
+    get_mesh_id,
+)
 
 
 def test_flow_match_scheduler_timesteps_monotone_decreasing() -> None:

tests/policies/lingbot_va/test_processor.py

@@ -21,6 +21,7 @@ import torch
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.policies.lingbot_va.configuration_lingbot_va import LingBotVAConfig
 from lerobot.policies.lingbot_va.processor_lingbot_va import (
+    LIBERO_ACTION_Q01,
     LingBotVAActionUnnormalizeStep,
     make_lingbot_va_pre_post_processors,
 )
@@ -75,7 +76,7 @@ def test_make_pre_post_processors_names_and_steps() -> None:
 def test_postprocessor_applies_unnormalization() -> None:
     cfg = _make_config()
     _, post = make_lingbot_va_pre_post_processors(cfg, dataset_stats=None)
-    # A normalized action of all -1 should map back to q01.
+    # A normalized action of all -1 should map back to q01 (the LIBERO 7-DoF default quantiles).
     normed = torch.full((1, len(cfg.used_action_channel_ids)), -1.0)
     out = post(normed)
-    assert torch.allclose(out, torch.tensor(cfg.action_q01).unsqueeze(0), atol=1e-4)
+    assert torch.allclose(out, torch.tensor(LIBERO_ACTION_Q01).unsqueeze(0), atol=1e-4)