cong-lab/stereo-ep-v1

Synchronized multi-camera stereo episodes

17 source recordings = 17 LeRobot episodes. Each row is one shared source timestamp, with all validated camera angles in camera-named observation fields. The same recording is never split into separate episodes by camera. There are 313 shared timestamps and 601 stereo camera observations across 6 compatible dataset roots.

Dataset root	Episodes	Cameras per episode
bench_vortex	4	front_zed_x_mini_stereo
box_test_tactile/agent_front	2	agent_stereo, front_stereo
box_test_tactile/agent_front_left	2	agent_stereo, front_stereo, left_stereo
pick_and_place	3	agent_stereo, front_stereo, left_stereo
pick_and_place_failed	3	agent_stereo, front_stereo, left_stereo
plate_spin	3	agent_stereo, front_stereo, left_stereo

The two box recordings with a rejected left view use the agent/front group. Other box recordings include agent/front/left. This keeps a fixed valid camera schema per LeRobot dataset without supplying missing camera images or calibration targets. Bench recordings have one validated camera; no additional validated angles currently exist for them.

Features at each timestamp

• observation.images.<camera>.left and .right: native source RGB eyes.
• observation.depth_m.<camera>: exact unfiltered float32 metric depth.
• observation.disparity_px.<camera>: exact float32 representation of stored disparity values.
• observation.confidence.<camera>: all original uint8 reliability scores.
• observation.cameras.<camera>.intrinsics, .source_eye_intrinsics, .camera_to_world, .world_to_camera, .baseline_m, .valid: that camera's calibration.
• Shared observation.state, source.episode_index, source.frame_index, source.timestamp_s and LeRobot episode/frame indices occur once per row.
• source.<camera>.*_path: paths to byte-exact raw files under modalities/episode-NNNNNN/<camera>/.

All cameras within an episode have identical source frame IDs, timestamps and robot state. Calibration maps the left optical camera to the C7 robot-base world frame, in metres. Intrinsics use a pixel-corner origin with pixel centres (u+0.5,v+0.5). Confidence is reliability, not a calibrated probability. Camera approval does not approve arm mesh fitting. No mesh targets or actions are synthesized.

Only previously approved complete camera streams and accepted geometry reference cameras are included;18 failed/partial camera streams remain excluded. Validation evidence and per-frame native calibration are retained. Bench intrinsics are legacy estimates; the other included rigs use recorded intrinsics.

Most recordings have12 sparse calibration timestamps; bench episode130 has121 exported timestamps (every third source frame). Original times are preserved; fps is the source clock and uniform_sampling=false. Use actual source timestamps for temporal sampling. Spatial-map statistics are global broadcastable summaries. No pixels are removed by a confidence cutoff, and no missing source frames or camera views are interpolated.

Load all camera angles together

from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

path = "pick_and_place"
root = snapshot_download("cong-lab/stereo-ep-v1", repo_type="dataset",
                         allow_patterns=[f"{path}/**"])
dataset = LeRobotDataset("cong-lab/stereo-ep-v1", root=Path(root) / path,
                         download_videos=False, video_backend="pyav")
frame = dataset[0]  # one timestamp, all three camera angles
front = frame["observation.images.front_stereo.left"]
agent = frame["observation.images.agent_stereo.left"]
left = frame["observation.images.left_stereo.left"]
depth = frame["observation.depth_m.front_stereo"]
T = frame["observation.cameras.front_stereo.camera_to_world"]

manifest.json lists source episode IDs and all included cameras. checksums.json covers the raw assets, Parquet files and metadata. The previous camera-separated layout remains available in Git history at 5c85b65074677452746f02311cd6ea144cefaa6a.

Per-step articulation estimates

Each episode now includes annotations/articulation/episode-NNNNNN.json. Six joint angles are fitted independently at each exported timestep, initialized from recorded joints and regularized toward their motion. One learned world_T_robot_base remains exactly constant across the entire episode. Camera poses, intrinsics, link geometry and raw pointmaps are fixed.

225/313 timesteps pass visible-arm validation. accepted_frame_mask is authoritative: failed fitted joint/link estimates are null, with no filling or interpolation. Recorded observation.state, RGB, stereo depth and confidence are unchanged. The results are geometric estimates rather than joint-encoder ground truth; unmodeled gripper and wrist hardware are excluded from the arm model. No mesh-derived track supervision is enabled by this annotation.

import json
pose_path = dataset.meta.info["articulation_path"].format(episode_index=0)
fit = json.loads((dataset.root / pose_path).read_text())
if fit["accepted_frame_mask"][0]:
    fitted_q = fit["q_fitted_rad"][0]  # six angles, radians
    episode_base = fit["world_T_robot_base"]  # same base for every timestep