basematrix/ego-binocular-v1

BaseMatrix EGO Binocular v1

Egocentric bimanual manipulation dataset with 3D hand tracking, EMG muscle signals, and robot-ready action representations. Captured from a first-person perspective using head-mounted stereo cameras and forearm EMG wristbands, processed through a 7-stage automated pipeline.

Key differentiators:

• Egocentric + binocular stereo — first-person view matching humanoid robot camera placement
• 21-joint 3D hand skeleton per hand (MANO topology) — retargetable to any robot hand
• 6-channel surface EMG (11/14 episodes) — muscle activation signals unavailable in any other public manipulation dataset
• Robot-agnostic — human demonstrations retarget to parallel-jaw grippers, dexterous hands, or humanoid manipulators via IK

Format: LeRobot v3.0 — load directly with HuggingFace datasets or LeRobotDataset.

Quick Start

from datasets import load_dataset

ds = load_dataset("basematrix/ego-binocular-v1", split="train")
print(f"{len(ds)} frames, {len(set(ds['episode_index']))} episodes")

frame = ds[1000]
left_kp3d = frame["observation.state.left_keypoints_3d"]   # 63D: 21 joints × 3
left_emg  = frame["observation.state.left_tactile"]         # 48D: 5 EMG features × 6ch + 18D reserved
right_wrist = frame["observation.state.right_wrist"]        # 3D: world-frame position (m)

Retarget to Your Robot (5 lines)

import numpy as np
from datasets import load_dataset

ds = load_dataset("basematrix/ego-binocular-v1", split="train")
ep3 = ds.filter(lambda x: x["episode_index"] == 3)

# For parallel-jaw grippers (Franka, UR5, ALOHA):
wrist_right = np.array(ep3["observation.state.right_wrist"])   # (T, 3) end-effector target
grasp_right = np.array(ep3["action.right_grasp"])              # (T,) 0=open, 1=closed
# → Feed wrist position to IK solver, grasp signal to gripper command

# For dexterous hands (LEAP, Allegro, Shadow, Ability):
kp3d = np.array(ep3["observation.state.right_keypoints_3d"])   # (T, 63)
kp3d = kp3d.reshape(-1, 21, 3)                                # (T, 21, 3) MANO topology
# → Use fingertip-position IK or our retarget_demo.py for joint angle mapping

A complete retargeting example with robot joint angle export is available at examples/retarget_demo.py.

Dataset Summary

Property	Value
Episodes	14
Total frames	77,511
Duration	~43 minutes
FPS	30
Scenes	Desktop manipulation (office environments)
Hands	Bimanual (left + right tracked independently)
EMG coverage	11/14 episodes (ep2–ep12, NB3 wristband, 99.3–100% either-hand, 89–97% per-hand)

Capture Hardware

Device	Specs	Mounting
Orbbec Femto Mega	Dual global-shutter fisheye, 1600×1300, 120mm stereo baseline, 1kHz IMU	Head-mounted (egocentric)
CEREBLINK NB3	6ch surface EMG @1kHz + 9-axis IMU @143Hz, BLE streaming	Forearm wristband
iPhone (optional)	1080p 30fps wide lens	Tripod (exocentric) — not included in this release

Features

Observation — Vision (2 video streams)

Feature	Shape	Description
observation.images.ego_left	(3, 1300, 1600)	Left stereo fisheye, AV1 30fps
observation.images.ego_right	(3, 1300, 1600)	Right stereo fisheye, AV1 30fps, frame-synchronized with ego_left

Observation — Hand State (9 features, ×2 hands)

Feature	Shape	Source	Description
observation.state.{side}_keypoints_3d	(63,)	WiLoR	21 MANO joints × 3D in a hand-local frame (meters); joint 0 sits ≈ 9.6 cm from the origin — subtract joint 0 for wrist-relative coordinates
observation.state.{side}_wrist	(3,)	HaWoR	Wrist position in world frame (meters)
observation.state.{side}_mano_pose	(45,)	HaWoR	MANO hand pose: 15 joints × axis-angle
observation.state.{side}_mano_rot	(3,)	HaWoR	MANO global orientation (axis-angle)
observation.state.{side}_mano_trans	(3,)	HaWoR	MANO translation in HaWoR's reconstruction frame — not the same frame as {side}_wrist; do not mix the two
observation.state.{side}_mano_betas	(10,)	WiLoR	MANO shape coefficients, estimated per frame (average over frames for a stable hand shape)
observation.state.{side}_contact	scalar	Hands23	Hand-object contact binary (0/1)
observation.state.{side}_tactile	(48,)	NB3 EMG	5 features × 6 channels = 30D active + 18D reserved
observation.state.{side}_force	(3,)	NB3 EMG	EMG-derived force proxy (accelerometer)

Observation — Scene State

Feature	Shape	Source	Description
observation.state.camera_pose	(7,)	ORB-SLAM3	6DoF camera pose: tx, ty, tz, qx, qy, qz, qw
observation.state.camera_pose_valid	(1,)	ORB-SLAM3	1 where the pose is backed by SLAM, 0 where it is not. Zero-pose frames must not be used as poses
observation.state.gripper_force	(2,)	NB3 EMG	EMG-derived grip force (left, right)

Action (6 features)

Feature	Shape	Description
action.{side}_wrist_delta	(3,)	Wrist displacement from the previous frame, wrist[t] − wrist[t−1] in meters (0 on an episode's first frame). This is a backward difference; for next-step targets use wrist[t+1] − wrist[t]
action.{side}_grasp	scalar	Grasp signal derived from contact + keypoint closure (0=open, 1=grasping)
action.{side}_gripper	(7,)	7D retarget vector from MANO skeleton: [thumb_aperture, index_curl, middle_curl, ring_curl, pinky_curl, spread, opposition]

Metadata

timestamp, frame_index, episode_index, index, task_index

EMG Muscle Signals

11 out of 14 episodes include 6-channel forearm EMG captured by CEREBLINK NB3 wristbands at 1kHz, downsampled and feature-extracted to 30Hz to align with video frames.

Each observation.state.{side}_tactile vector (48D) layout:

• Index 0–5: RMS × 6 channels
• Index 6–11: MAV (mean absolute value) × 6 channels
• Index 12–17: WL (waveform length) × 6 channels
• Index 18–23: ZC (zero-crossing rate) × 6 channels
• Index 24–29: SSC (slope sign changes) × 6 channels
• Index 30–47: reserved (zeros in current version)

Why EMG matters for robot learning:

• EMG activates 50–200ms before physical contact — a predictive signal for grasp timing
• Co-contraction patterns encode intended stiffness/compliance
• Force intention without requiring robot-mounted sensors

To select EMG-active frames:

import numpy as np
emg = np.array(ds["observation.state.left_tactile"])
emg_active = np.abs(emg).sum(axis=-1) > 0  # True for episodes with wristband data

Hand Tracking Pipeline

Three complementary vision models provide layered hand understanding:

Model	Output	Role
Hands23 (HOI-DETR)	Bounding boxes, contact state, grasp type	When and what is being grasped
WiLoR-mini	21 MANO keypoints in 3D (per-frame)	Hand shape and finger articulation
HaWoR	Wrist trajectories in world coordinates + MANO pose/orientation/translation (temporal)	Globally consistent hand position and articulation

The 21-keypoint skeleton follows MANO topology:

• Joint 0: wrist
• Joints 1–4: thumb (CMC → MCP → IP → tip)
• Joints 5–8: index (MCP → PIP → DIP → tip)
• Joints 9–12: middle, 13–16: ring, 17–20: little

Full Processing Pipeline

Stage	Tool	Output
L0	Audio beep detection (1kHz threshold)	Multi-camera time alignment
L1	Nearest-neighbor PTS matching	Stereo frame pairs + NB3 EMG sync
L2	ORB-SLAM3 (stereo)	6DoF camera trajectory
L3	Hands23 (HOI-DETR)	Hand-object contact + grasp classification
L3b	WiLoR-mini	Single-frame 3D hand skeleton
L3c	HaWoR	Temporal wrist trajectory in world frame
L4	GPT-4.1 Vision	Semantic annotation (objects, phases)
L5	Cross-layer verification (24 checks)	Quality score + data card
L6	LeRobot v3.0 packaging	This dataset

Data Quality

Metric	Value	Notes
NaN/Inf	0%	All numeric features verified clean
Camera pose coverage	100%	77,505 of 77,511 frames carry a SLAM-backed pose (camera_pose_valid == 1); the six exceptions sit past the last keyframe of their episode
Keypoint zero frames	L: 8.7%, R: 5.3%	Hand outside FOV or WiLoR detection failure; use abs(kp3d).sum() > 0 as validity mask
Wrist trajectory outliers	L: 1.5%, R: 2.3%	HaWoR tracking loss → delta > 3 m/s; apply median filter or clip to ≤0.1 m/frame
EMG frame coverage	99.3–100% either-hand	Per-hand: L 89–97%, R 93–96%. At least one hand has EMG in 99.3–100% of frames
Contact detection rate	97.6%	Hands23 successfully classified contact in 97.6% of hand-visible frames (mean across episodes; range 92.3–99.4%)

Coordinate System

• World frame: right-handed, Y-up; origin at ORB-SLAM3 initialization point. Scale is metric, recovered from the 120 mm stereo baseline rather than from the IMU
• Keypoints (63D): hand-local frame in meters (WiLoR); subtract joint 0 for wrist-relative coordinates
• Wrist position (3D): absolute world-frame coordinates in meters (HaWoR)
• MANO translation (3D): HaWoR reconstruction frame — differs from the wrist-position frame
• MANO parameters: zero on frames where hand tracking was lost
• Camera pose (7D): position (tx, ty, tz) + unit quaternion (qx, qy, qz, qw)

Use Cases

Application	Which features to use
Imitation learning (ACT, Diffusion Policy)	ego_left + {side}_wrist + {side}_grasp
VLA pre-training (π0, Octo, RT-2)	All vision + state + task descriptions
Dexterous hand retargeting	{side}_keypoints_3d (63D) → IK to target hand
Parallel-jaw gripper control	{side}_wrist_delta + {side}_grasp
Grasp timing prediction	{side}_tactile (EMG) + {side}_contact
Force-aware manipulation	{side}_tactile + gripper_force

Comparison with Teleoperation Datasets

Property	This dataset (EGO)	Typical teleoperation
Robot-specific?	No — retarget to any robot	Yes — locked to source robot
Hand representation	21-joint 3D skeleton (63D)	1D gripper open/close
Force sensing	EMG muscle signals (predictive)	F/T sensor (reactive)
Perspective	Egocentric (matches humanoid cameras)	Third-person fixed cameras
Collection cost	Low (no robot hardware needed)	High (robot + operator)
Scaling	Parallelizable (multiple wearers)	Sequential (one robot at a time)

Roadmap

• Scale to 50+ episodes with diverse household manipulation tasks
• Add multi-environment recordings (kitchen, workshop, warehouse)

Changelog

2026-09-24 — camera trajectory fix, quality gate applied (this revision)

• Episode 0's camera trajectory was wrong and is now corrected. Its SLAM input frames had been extracted at full resolution (1600×1300) while the camera intrinsics accompanying them were written for the half-resolution setting the pipeline uses everywhere else. Image and intrinsics disagreed, so the reconstruction came out mis-scaled: the episode reported 10.56 m of camera travel where the correct figure is 2.06 m, with a correspondingly inflated bounding box. camera_pose for episode 0 has been re-derived from self-consistent inputs. The other 13 episodes were already consistent and their poses are unchanged.
• One episode removed, one added. The Aug-17 recording (previously episode 2, and the only one that carried an exocentric view) does not pass the pipeline's image-sharpness gate — 9 of 10 sampled left-camera frames are blurred — and is no longer part of the dataset. A Sep-21 recording ("organizing desk electronics and cables", 8,871 frames) is included in its place.
• Totals changed accordingly: still 14 episodes, but 70,754 → 77,511 frames (~39 → ~43 minutes). Episode indices after the removed recording have shifted by one.
• EMG episodes are now ep2–ep12 (previously ep3–ep13), still 11 of 14. If you select EMG episodes by index, update the range.
• Data-quality figures in the table above were recomputed against this episode set (pose coverage, keypoint zero frames, wrist outliers, contact rate).

2026-09-13 — re-packaged from source

The previous revision (f03d1f6c) had packaging defects. If you downloaded it, please download again.

• Stereo sync fixed. Videos were the raw camera files starting at their own first frame, while each row is defined by a synchronized left/right frame pair. The right view was 1–19 frames out of sync in 13 of 14 episodes (episode 5: the left view, by 12 frames). Each view is now re-encoded so that video frame i is exactly the frame row i refers to.
• Loads at LeRobot's default timestamp tolerance. Video timestamps now start at 0 at a constant 30 fps, matching the timestamp column. Previously most frames failed with FrameTimestampError.
• Hand data realigned in episode 5. Wrist, MANO and keypoint columns were offset by 12 frames against contact, camera pose and EMG.
• MANO columns now present. The previous meta/info.json declared the 8 MANO columns, but the parquet did not contain them.
• 8 trailing rows removed (70,762 → 70,754). The camera leaves a truncated final packet with no image; rows whose frame fell on it were dropped (episodes 3–7, 9, 12, 13).
• EXO camera removed. It existed for 1 of 14 episodes, and declaring it made loading fail for the other 13.
• Image normalization statistics added to meta/stats.json. Videos are encoded as AV1 (GOP 2), LeRobot's default.

Citation

@misc{basematrix2026ego,
  title={BaseMatrix EGO Binocular v1: Egocentric Hand Manipulation Dataset with EMG for Robot Learning},
  author={BaseMatrix},
  year={2026},
  url={https://huggingface.co/datasets/basematrix/ego-binocular-v1}
}

License

CC BY-NC 4.0 — free for research and evaluation. Contact info@basematrix.ai for commercial licensing.

Links

• Retarget demo: examples/retarget_demo.py — MANO keypoints → robot joint angles, export to CSV
• Website: basematrix.ai
• X/Twitter: @BaseMatrixAI
• Email: info@basematrix.ai