--- license: apache-2.0 task_categories: - robotics tags: - LeRobot - robotics - mujoco - robosuite - manipulation - event-camera - dvs configs: - config_name: default data_files: data/*/*.parquet --- This dataset was created using [LeRobot](https://github.com/huggingface/lerobot). ## Dataset Description # dvla-can-250hz-events (250 fps native) - event frames accumulating 4 ms each 1016 episodes / 1,356,080 frames at the native 250 Hz rate. Same v2e settings as the 250->25 fps variant, but kept at full rate: each event frame covers 4 ms, so the polarity images are sparser (roughly a third of the pixels of the 40 ms version). Two extra video columns on top of the 3 RGB cameras: `observation.images.opst_cam_events`, `observation.images.wrist_cam_events`. Note on encoding: SVT-AV1 refuses frame rates above 240 fps, so every video column here is H.264 rather than AV1. The 25 fps variants use AV1 for the base and H.264 after the trim re-encode. The 250 Hz RGB counterpart of this dataset trains to only **36%** place success versus 61% at 25 Hz, and the deficit is entirely on *moving* cans (static cans stay at 100%). This dataset exists to test whether event input recovers that gap. ### Task MuJoCo / robosuite (Panda + OSC-or-IK) `place`: pick a can off the table and drop it into a bowl. In 80% of the episodes the can is **rolling** when the episode starts (speed sampled from 0.25-0.75 m/s); the rest are static. Demonstrations come from a scripted state machine that reads privileged simulator state, and only successful attempts are kept. ### Layout LeRobot v2.1, `robot_type: panda`, 3 RGB cameras at 360x480 (`wrist_cam`, `side_cam`, `opst_cam`). Training uses `opst_cam` + `wrist_cam` only. - `action` (10,) = `[dx, dy, dz, sin/cos of the 3 euler angles, gripper]` - `observation.state` (9,) = `[x, y, z, sin/cos of the 3 euler angles]` - `observation.environment_state` (9,) = privileged object pose/velocity. Dropped during training (it is not available on a real robot). `action` is a **delta**: the absolute target minus the state of that same frame. At evaluation the policy output is decoded as `target = live_state + predicted_delta`, re-anchored every step. The absolute target it encodes is a **future-EE relabel**: instead of the pose the state machine commanded at time t, the label is the pose the arm had actually *reached* 320 ms later. A sweep over that offset found it matters enormously - replaying the labels open-loop succeeds 20/20 at 320 ms but 5/20 at 200 ms and 0/20 at 40 ms. `-delta-trim` additionally drops the leading frames where the arm is still holding position, so that `action[0]` is a real motion rather than a placeholder (which otherwise causes a cold-start stall). `meta/camera.jsonl` maps each `episode_index` back to its source HDF5 filename. Conversion shards round-robin, so episode order is **not** source order and this file is the only way back. ### Raw event streams (`raw_events/`) The event video columns are **rasterised** frames: where several events land on the same pixel only one survives, discarding 72-73% of them. The stream before that rasterisation ships in this repo under `raw_events/` (1016 files, 20 GB). Two arrays per camera: ``` opst_cam_events_xytp (N, 4) int32 one event per row: [x, y, t_us, polarity] opst_cam_events_idx (T+1,) int32 frame boundaries wrist_cam_events_xytp (M, 4) int32 wrist_cam_events_idx (T+1,) int32 ``` `x` 0-479, `y` 0-359, polarity **+1** (brighter) / **-1** (darker), sorted by time. Events of frame `i` are `xytp[idx[i] : idx[i+1]]`; `idx[-1]` equals the total event count. Roughly 1.3-1.7 M events per episode for `opst_cam`, 4-8 M for `wrist_cam`. ```python import json, h5py cam = {json.loads(l)["episode_index"]: json.loads(l)["filename"] for l in open(f"{root}/meta/camera.jsonl")} # episode -> source filename with h5py.File(f"{root}/raw_events/{cam[ep]}", "r") as h: idx = h["opst_cam_events_idx"][:] # small, read whole s, e = int(idx[frame]), int(idx[frame + span]) # span frames = span x 4 ms ev = h["opst_cam_events_xytp"][s:e] # only this slice leaves disk ``` Episode order is not source order (conversion shards round-robin), so `meta/camera.jsonl` is the only way to pair an episode with its raw stream. With these you can build other representations -- event counts, time surfaces, voxel grids, other accumulation windows -- **without re-running v2e**, which is the most expensive stage of the pipeline (5.8 h for 1016 episodes). > **Training only? Skip this folder.** Opening `LeRobotDataset(repo_id)` without an `episodes` > argument calls `snapshot_download` with no allow-list, so all 20 GB come along: > ```python > from huggingface_hub import snapshot_download > snapshot_download(repo_id, repo_type="dataset", local_dir=root, ignore_patterns="raw_events/") > LeRobotDataset(repo_id, root=root) > ``` ### In-domain eval scenes (`eval_scenes/`) Reproducing the in-domain numbers needs the scenes restored exactly as they were generated, which `run_ckpt_eval*.sh` does through `--replay_h5_dir` on the source HDF5s. Those are 37 GB (25 Hz) / 267 GB (250 Hz) and are not published. They do not have to be. `eval_server.load_h5_episode` reads only the root attrs (`env_config`, `seed`, `instruction`, `init_*`) and the `action` dataset -- the camera frames, which are 99.99% of the file, are re-rendered by the server. Stripped to that, a 250 Hz 20-scene set goes from 5.1 GB to 0.37 MB, and ships here under `eval_scenes/`. Restoring from the stripped files was verified to reproduce the original scene exactly: object pose and velocity, bowl pose, arm pose and action length all match. ```bash huggingface-cli download mickeykang/dvla-can-250hz-events-250fps-delta-trim --repo-type dataset --include 'eval_scenes/*' --local-dir . REPLAY_DIR=$PWD/eval_scenes/indomain_scenes_s20_250 ... bash overnight_delta_n3/run_policy_eval.sh ``` Out-of-distribution scenes ship here too, under `eval_scenes/ood_scenes_s20_250`. They are normally sampled from seeds 9000000-9000019 at run time, but that only reproduces on the machine that produced the published numbers: the sampler's bowl-clearance test reads mesh radii out of `assets/`, and the result also depends on the robosuite/MuJoCo build and the floating-point path. Replaying the frozen scenes removes all of that. The dumped scene names match the ones in the original eval logs, and replaying reproduces the seed-sampled scene exactly (object, bowl and joints identical; end-effector differs by 3e-08 because `init_*` is stored float32, a property of the original pipeline). ```bash REPLAY_DIR=$PWD/eval_scenes/ood_scenes_s20_250 ... bash overnight_delta_n3/run_policy_eval.sh ``` ### Reproducing Generation, conversion, training and evaluation scripts, plus the measured results and the gotchas that cost the most time, are documented at https://github.com/mickeykang16/DynamicVLA/tree/mujoco. - **Homepage:** https://github.com/mickeykang16/DynamicVLA/tree/mujoco - **Paper:** [More Information Needed] - **License:** apache-2.0 ## Dataset Structure [meta/info.json](meta/info.json): ```json { "codebase_version": "v2.1", "robot_type": "panda", "total_episodes": 1016, "total_frames": 1356080, "total_tasks": 3, "total_videos": 5080, "total_chunks": 2, "chunks_size": 1000, "fps": 250, "splits": { "train": "0:1016" }, "data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet", "video_path": "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4", "features": { "action": { "dtype": "float32", "shape": [ 10 ], "names": [ "ee_pos_x", "ee_pos_y", "ee_pos_z", "ee_rot_x_sin", "ee_rot_x_cos", "ee_rot_y_sin", "ee_rot_y_cos", "ee_rot_z_sin", "ee_rot_z_cos", "gripper" ] }, "observation.state": { "dtype": "float32", "shape": [ 9 ], "names": [ "ee_pos_x", "ee_pos_y", "ee_pos_z", "ee_rot_x_sin", "ee_rot_x_cos", "ee_rot_y_sin", "ee_rot_y_cos", "ee_rot_z_sin", "ee_rot_z_cos" ] }, "observation.environment_state": { "dtype": "float32", "shape": [ 9 ], "names": [ "o0", "o1", "o2", "o3", "o4", "o5", "o6", "o7", "o8" ] }, "observation.images.wrist_cam": { "dtype": "video", "shape": [ 360, 480, 3 ], "names": [ "height", "width", "channel" ], "info": { "video.height": 360, "video.width": 480, "video.codec": "h264", "video.pix_fmt": "yuv420p", "video.is_depth_map": false, "video.fps": 250, "video.channels": 3, "has_audio": false } }, "observation.images.side_cam": { "dtype": "video", "shape": [ 360, 480, 3 ], "names": [ "height", "width", "channel" ], "info": { "video.height": 360, "video.width": 480, "video.codec": "h264", "video.pix_fmt": "yuv420p", "video.is_depth_map": false, "video.fps": 250, "video.channels": 3, "has_audio": false } }, "observation.images.opst_cam": { "dtype": "video", "shape": [ 360, 480, 3 ], "names": [ "height", "width", "channel" ], "info": { "video.height": 360, "video.width": 480, "video.codec": "h264", "video.pix_fmt": "yuv420p", "video.is_depth_map": false, "video.fps": 250, "video.channels": 3, "has_audio": false } }, "observation.images.wrist_cam_events": { "dtype": "video", "shape": [ 360, 480, 3 ], "names": [ "height", "width", "channel" ], "info": { "video.height": 360, "video.width": 480, "video.codec": "h264", "video.pix_fmt": "yuv420p", "video.is_depth_map": false, "video.fps": 250, "video.channels": 3, "has_audio": false } }, "observation.images.opst_cam_events": { "dtype": "video", "shape": [ 360, 480, 3 ], "names": [ "height", "width", "channel" ], "info": { "video.height": 360, "video.width": 480, "video.codec": "h264", "video.pix_fmt": "yuv420p", "video.is_depth_map": false, "video.fps": 250, "video.channels": 3, "has_audio": false } }, "timestamp": { "dtype": "float32", "shape": [ 1 ], "names": null }, "frame_index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "episode_index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "task_index": { "dtype": "int64", "shape": [ 1 ], "names": null } } } ``` ## Citation **BibTeX:** ```bibtex [More Information Needed] ```