Merge remote-tracking branch 'origin/main' into codex/episode-video-streaming-byte-cache

src/lerobot/datasets/lerobot_dataset.py

@@ -478,18 +478,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """Return the number of frames in the selected episodes."""
         return self.num_frames
 
-    def __getitem__(self, idx) -> dict:
-        """Return a single frame by index, with all transforms applied.
+    def __getitem__(self, idx: int | slice) -> dict | list[dict]:
+        """Return one frame or a slice of frames, with all transforms applied.
 
         Loads the frame from the underlying HF dataset, expands delta-timestamp
         windows, decodes video frames, and applies image transforms. Delegates
-        the core logic to :meth:`DatasetReader.get_item`.
+        the core logic to :class:`DatasetReader`.
 
         Args:
-            idx: Index into the (possibly episode-filtered) dataset.
+            idx: Integer index or slice into the possibly episode-filtered dataset.
 
         Returns:
-            Dict mapping feature names to their tensor values for this frame.
+            A frame dictionary for an integer index, or a list of frame
+            dictionaries for a slice.
 
         Raises:
             RuntimeError: If the dataset is currently being recorded and
@@ -499,6 +500,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
             raise RuntimeError(
                 "Cannot read from a dataset that is being recorded. Call finalize() first, then access items."
             )
+        if isinstance(idx, slice):
+            return [self[item_idx] for item_idx in range(*idx.indices(len(self)))]
+
         reader = self._ensure_reader()
         if reader.hf_dataset is None:
             # One-shot load after finalize()

src/lerobot/envs/configs.py

@@ -322,7 +322,7 @@ class HILSerlRobotEnvConfig(EnvConfig):
 class LiberoEnv(EnvConfig):
     task: str = "libero_10"  # can also choose libero_spatial, libero_object, etc.
     task_ids: list[int] | None = None
-    fps: int = 30
+    fps: int = 20  # Must match robosuite's default control_freq (20 Hz)
     episode_length: int | None = None
     obs_type: str = "pixels_agent_pos"
     render_mode: str = "rgb_array"
@@ -354,6 +354,9 @@ class LiberoEnv(EnvConfig):
     control_mode: str = "relative"  # or "absolute"
 
     def __post_init__(self):
+        if self.fps <= 0:
+            raise ValueError(f"fps must be positive, got {self.fps}")
+
         if self.obs_type == "pixels":
             self.features[LIBERO_KEY_PIXELS_AGENTVIEW] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
@@ -412,6 +415,7 @@ class LiberoEnv(EnvConfig):
             "render_mode": self.render_mode,
             "observation_height": self.observation_height,
             "observation_width": self.observation_width,
+            "control_freq": self.fps,
         }
         if self.task_ids is not None:
             kwargs["task_ids"] = self.task_ids

src/lerobot/envs/libero.py

@@ -125,10 +125,13 @@ class LiberoEnv(gym.Env):
         n_envs: int = 1,
         camera_name_mapping: dict[str, str] | None = None,
         num_steps_wait: int = 10,
+        control_freq: int = 20,
         control_mode: str = "relative",
         is_libero_plus: bool = False,
     ):
         super().__init__()
+        if control_freq <= 0:
+            raise ValueError(f"control_freq must be positive, got {control_freq}")
         self.task_id = task_id
         self.is_libero_plus = is_libero_plus
         self.obs_type = obs_type
@@ -154,6 +157,7 @@ class LiberoEnv(gym.Env):
             }
         self.camera_name_mapping = camera_name_mapping
         self.num_steps_wait = num_steps_wait
+        self.control_freq = control_freq
         self.episode_index = episode_index
         self.episode_length = episode_length
         # Load once and keep
@@ -260,6 +264,7 @@ class LiberoEnv(gym.Env):
             bddl_file_name=self._task_bddl_file,
             camera_heights=self.observation_height,
             camera_widths=self.observation_width,
+            control_freq=self.control_freq,
         )
         env.reset()
         self._env = env

src/lerobot/motors/damiao/damiao.py

@@ -20,7 +20,6 @@ import logging
 import time
 from contextlib import contextmanager
 from copy import deepcopy
-from functools import cached_property
 from typing import TYPE_CHECKING, Any, TypedDict
 
 from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
@@ -854,7 +853,7 @@ class DamiaoMotorsBus(MotorsBusBase):
         else:
             raise ValueError(f"Motor {motor_obj} doesn't have a valid recv_id (None).")
 
-    @cached_property
+    @property
     def is_calibrated(self) -> bool:
         """Check if motors are calibrated."""
         return bool(self.calibration)

src/lerobot/motors/motors_bus.py

@@ -23,6 +23,7 @@ from __future__ import annotations
 
 import abc
 import logging
+import time
 from collections.abc import Sequence
 from contextlib import contextmanager
 from dataclasses import dataclass
@@ -818,13 +819,13 @@ class SerialMotorsBus(MotorsBusBase):
         """
         motor_names = self._get_motors_list(motors)
 
-        start_positions = self.sync_read("Present_Position", motor_names, normalize=False)
+        start_positions = self.sync_read("Present_Position", motor_names, normalize=False, num_retry=5)
         mins = start_positions.copy()
         maxes = start_positions.copy()
 
         user_pressed_enter = False
         while not user_pressed_enter:
-            positions = self.sync_read("Present_Position", motor_names, normalize=False)
+            positions = self.sync_read("Present_Position", motor_names, normalize=False, num_retry=5)
             mins = {motor: min(positions[motor], min_) for motor, min_ in mins.items()}
             maxes = {motor: max(positions[motor], max_) for motor, max_ in maxes.items()}
 
@@ -837,9 +838,12 @@ class SerialMotorsBus(MotorsBusBase):
             if enter_pressed():
                 user_pressed_enter = True
 
-            if display_values and not user_pressed_enter:
-                # Move cursor up to overwrite the previous output
-                move_cursor_up(len(motor_names) + 3)
+            if not user_pressed_enter:
+                if display_values:
+                    # Move cursor up to overwrite the previous output
+                    move_cursor_up(len(motor_names) + 3)
+                # Throttle reads even when the live table is disabled.
+                time.sleep(0.02)
 
         same_min_max = [motor for motor in motor_names if mins[motor] == maxes[motor]]
         if same_min_max:

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -79,6 +79,8 @@ class DiffusionConfig(PreTrainedConfig):
         use_film_scale_modulation: FiLM (https://huggingface.co/papers/1709.07871) is used for the Unet conditioning.
             Bias modulation is used be default, while this parameter indicates whether to also use scale
             modulation.
+        gradient_checkpointing: Whether to checkpoint the Unet residual blocks during training. This reduces
+            activation memory at the cost of recomputing those blocks during the backward pass.
         noise_scheduler_type: Name of the noise scheduler to use. Supported options: ["DDPM", "DDIM"].
         num_train_timesteps: Number of diffusion steps for the forward diffusion schedule.
         beta_schedule: Name of the diffusion beta schedule as per DDPMScheduler from Hugging Face diffusers.
@@ -132,6 +134,7 @@ class DiffusionConfig(PreTrainedConfig):
     n_groups: int = 8
     diffusion_step_embed_dim: int = 128
     use_film_scale_modulation: bool = True
+    gradient_checkpointing: bool = False
     # Noise scheduler.
     noise_scheduler_type: str = "DDPM"
     num_train_timesteps: int = 100

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -31,6 +31,7 @@ import torch
 import torch.nn.functional as F  # noqa: N812
 import torchvision
 from torch import Tensor, nn
+from torch.utils.checkpoint import checkpoint
 
 from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
 from lerobot.utils.import_utils import _diffusers_available, require_package
@@ -727,22 +728,35 @@ class DiffusionConditionalUnet1d(nn.Module):
         else:
             global_feature = timesteps_embed
 
+        use_gc = self.config.gradient_checkpointing and self.training
+
         # Run encoder, keeping track of skip features to pass to the decoder.
         encoder_skip_features: list[Tensor] = []
         for resnet, resnet2, downsample in self.down_modules:
-            x = resnet(x, global_feature)
-            x = resnet2(x, global_feature)
+            if use_gc:
+                x = checkpoint(resnet, x, global_feature, use_reentrant=False)
+                x = checkpoint(resnet2, x, global_feature, use_reentrant=False)
+            else:
+                x = resnet(x, global_feature)
+                x = resnet2(x, global_feature)
             encoder_skip_features.append(x)
             x = downsample(x)
 
         for mid_module in self.mid_modules:
-            x = mid_module(x, global_feature)
+            if use_gc:
+                x = checkpoint(mid_module, x, global_feature, use_reentrant=False)
+            else:
+                x = mid_module(x, global_feature)
 
         # Run decoder, using the skip features from the encoder.
         for resnet, resnet2, upsample in self.up_modules:
             x = torch.cat((x, encoder_skip_features.pop()), dim=1)
-            x = resnet(x, global_feature)
-            x = resnet2(x, global_feature)
+            if use_gc:
+                x = checkpoint(resnet, x, global_feature, use_reentrant=False)
+                x = checkpoint(resnet2, x, global_feature, use_reentrant=False)
+            else:
+                x = resnet(x, global_feature)
+                x = resnet2(x, global_feature)
             x = upsample(x)
 
         x = self.final_conv(x)

src/lerobot/robots/openarm_follower/openarm_follower.py

@@ -150,9 +150,6 @@ class OpenArmFollower(Robot):
 
         self.configure()
 
-        if self.is_calibrated:
-            self.bus.set_zero_position()
-
         self.bus.enable_torque()
 
         logger.info(f"{self} connected.")

src/lerobot/scripts/lerobot_setup_motors.py

@@ -51,19 +51,7 @@ from lerobot.teleoperators import (  # noqa: F401
     rebot_102_leader,
     so_leader,
 )
-
-COMPATIBLE_DEVICES = [
-    "koch_follower",
-    "koch_leader",
-    "omx_follower",
-    "omx_leader",
-    "openarm_mini",
-    "so100_follower",
-    "so100_leader",
-    "so101_follower",
-    "so101_leader",
-    "lekiwi",
-]
+from lerobot.utils.import_utils import register_third_party_plugins
 
 
 @dataclass
@@ -80,18 +68,19 @@ class SetupConfig:
 
 @draccus.wrap()
 def setup_motors(cfg: SetupConfig):
-    if cfg.device.type not in COMPATIBLE_DEVICES:
-        raise NotImplementedError
-
     if isinstance(cfg.device, RobotConfig):
         device = make_robot_from_config(cfg.device)
     else:
         device = make_teleoperator_from_config(cfg.device)
 
-    device.setup_motors()
+    setup = getattr(device, "setup_motors", None)
+    if not callable(setup):
+        raise NotImplementedError(f"Device type '{cfg.device.type}' does not support motor setup.")
+    setup()
 
 
 def main():
+    register_third_party_plugins()
     setup_motors()
 
 

src/lerobot/teleoperators/gamepad/configuration_gamepad.py

@@ -23,3 +23,5 @@ from ..config import TeleoperatorConfig
 @dataclass
 class GamepadTeleopConfig(TeleoperatorConfig):
     use_gripper: bool = True
+    # Use hidapi instead of pygame for controllers that pygame cannot detect reliably.
+    hidapi_fallback: bool = False

src/lerobot/teleoperators/gamepad/teleop_gamepad.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import logging
 import sys
 from enum import IntEnum
 from typing import Any
@@ -27,6 +28,8 @@ from ..teleoperator import Teleoperator
 from ..utils import TeleopEvents
 from .configuration_gamepad import GamepadTeleopConfig
 
+logger = logging.getLogger(__name__)
+
 
 class GripperAction(IntEnum):
     CLOSE = 0
@@ -56,6 +59,13 @@ class GamepadTeleop(Teleoperator):
 
         self.gamepad = None
 
+        self.hidapi_fallback = config.hidapi_fallback
+        if sys.platform == "darwin" and not self.hidapi_fallback:
+            logger.warning(
+                "On macOS, pygame may not reliably detect input from some controllers. "
+                "If you experience issues, set `hidapi_fallback=true`."
+            )
+
     @property
     def action_features(self) -> dict:
         if self.config.use_gripper:
@@ -76,9 +86,7 @@ class GamepadTeleop(Teleoperator):
         return {}
 
     def connect(self) -> None:
-        # use HidApi for macos
-        if sys.platform == "darwin":
-            # NOTE: On macOS, pygame doesn’t reliably detect input from some controllers so we fall back to hidapi
+        if self.hidapi_fallback:
             from .gamepad_utils import GamepadControllerHID as Gamepad
         else:
             from .gamepad_utils import GamepadController as Gamepad