EgoVerse Aria ego subset, retrieved with real-robot trajectories (238 episodes)
Human egocentric demonstrations (Meta Aria glasses, EgoVerse release) selected for co-training a bimanual
parallel-gripper policy together with Noctis12138/rokae_bimanual_6task.
Format: LeRobot v2.1 (what openpi's pinned lerobot reads). Size: 238 episodes,
1,145,600 frames (10.6 h) at 30 Hz, 7.7 GB.
Private. Do not redistribute. The EgoVerse data terms (including partner-sourced parts) have not been checked for redistribution; this copy is for internal training only.
How the 238 episodes were chosen
Funnel over the EgoVerse Aria pool (1,882 episodes): quality filter Q35 ∩ L2 (Q35 = best 35% by quality within each
task cluster), then L3 task relevance P50 = the global top 50% of all ego episodes ranked by similarity of
motion descriptors (arc-length-normalised speed profile, curvature, height; rank fusion F2) to the wrist trajectories of
all 6 real-robot tasks in the companion dataset (the "r6" list). No per-task quota at the relevance step.
Then all fold-clothes episodes were removed (source task_name fold_clothes and fold_clothes_indomain, 64 of the 302 P50 episodes) by decision of the project lead, who judged folding too far from the robot tasks. At the relevance step fold clothes also had the lowest pass rate of all families (26% vs 53-100%).
208 of these 238 episodes are also in the earlier 312-episode list retrieved with 2 robot tasks. Selection used
trajectories only, not images or text; "similar" means similar trajectory shape, not similar task content.
| task_name (source) | episodes |
|---|---|
| cup_on_saucer | 83 |
| bag_groceries | 46 |
| pick_place | 36 |
| cup_on_saucer_static_cam | 22 |
| sort | 17 |
| put cup on saucer indomain | 11 |
| bag_groceries_success_2 | 8 |
| cup_on_saucer_in_domain | 6 |
| bag_groceries_success_3 | 5 |
| EgoWAM_bag_groceries | 2 |
| EgoWAM_cup_on_saucer | 1 |
| sort_test | 1 |
Prompts (task, from the source task_description):
| prompt | episodes |
|---|---|
| bagging groceries | 38 |
| put cup on saucer | 34 |
| putting cup on saucer | 33 |
| pick up cup with one arm, hand it over to the other arm and place it on the saucer | 27 |
| sort base objects | 19 |
| set 2 base data | 15 |
| pick up the bag and make it upright, place the objects into the bag | 15 |
| pick up cup with one arm, hand it over to the other arm and place it on saucer | 13 |
| set 1 base data | 12 |
| pick cup with one arm, hand it to the other arm and place it on the saucer | 7 |
| put 3 objects into a grocery bag | 7 |
| put cup on saucer in domain human data 2 hours | 3 |
| set 2 base data color correct | 3 |
| set 1 base data recollect color correct | 3 |
| In this task, the collector will sit or stand in front of a table that has a cup on the left/right, and a saucer on the other side of the table. The collector will use one hand to pick up the cup, hand over the cup to the other hand, and place the cup on the saucer. | 2 |
| placing cup on sacuer | 2 |
| put groceries into a bag | 1 |
| placing cup on saucer | 1 |
| put cup in saucer | 1 |
| sort base dataset object test data | 1 |
| set 3 base data | 1 |
# openpi pins lerobot 0cf8648 (codebase v2.1); the repo carries the `v2.1` tag the loader asks for.
# Private repo: `huggingface-cli login` (or HF_TOKEN) with an account that has access.
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("Noctis12138/egoverse-aria-selected", delta_timestamps={"action": [t / 30 for t in range(50)]})
item = ds[0] # item["action"]: (50, 16) absolute EE poses; item["task"]: prompt string
Features
| key | shape | content |
|---|---|---|
| observation.images.head | 480×640×3 video (AV1) | Aria front camera front_1, rectified pinhole, fx = fy = 266.5, cx = 320, cy = 240 |
| observation.state | 16 | hand EE poses, layout below |
| action | 16 | identical to observation.state (human data has no commands) |
| loss_mask | 16 | 1 = supervise; gripper dims (7, 15) are always 0; a hand's 7 pose dims are 0 where it is untracked |
| hand_tracked | 2 | 1 where the hand was tracked in the source (left, right) |
| hand_interpolated | 2 | 1 where a short gap (≤ 15 frames = 0.5 s) was filled by interpolation |
| frame_valid | 1 | 0 where a used hand (tracked in ≥ 50% of its episode) is in a longer gap; drop windows touching these frames |
| observation.raw.ee_pose_mps | 14 | source obs_ee_pose per hand: xyz + 4 quaternion numbers in stored order (order not verified; likely w,x,y,z) |
| observation.raw.wrist_pose_mps | 14 | source obs_wrist_pose per hand, same convention |
| observation.raw.head_pose | 7 | source obs_head_pose |
| observation.raw.hand_keypoints | 126 | 21 keypoints × xyz per hand (left then right), MediaPipe order (0 wrist, 4 thumb tip, 8 index tip), world frame |
16-d EE layout
| dims | meaning |
|---|---|
| 0-2 | left EE position x, y, z (m) |
| 3-6 | left EE orientation quaternion (x, y, z, w), sign made continuous over time |
| 7 | left gripper |
| 8-10 | right EE position x, y, z (m) |
| 11-14 | right EE orientation quaternion (x, y, z, w) |
| 15 | right gripper |
Definitions (human side)
- Frame: the EgoVerse/Aria world frame (z up, gravity-aligned). Both hands share it.
- Position: the source
obs_ee_poseposition, which sits at the palm centre (1.2 cm from the Aria palm keypoint). - Orientation, rebuilt from keypoints so that it follows a parallel-gripper convention: z = wrist (kp0) → midpoint of thumb tip (kp4) and index tip (kp8); y = thumb tip → index tip, made orthogonal to z; x = y × z. When the part of thumb→index orthogonal to z is < 1 cm (fingertips pinched together), the previous frame's y is reused. This happened on 6.8% of hand-frames (mean over episodes).
- Gripper slot: thumb-tip to index-tip distance in metres. Informational only;
loss_maskis 0 there because the plan is to learn the gripper from robot data only. - Untracked hand = all-zero row, NaN, all-zero keypoints, > 1 m from the head, or a row frozen at the previous value. Masked frames are filled with the nearest good value so every array is finite.
Tracking statistics (from the conversion)
| left | right | |
|---|---|---|
| tracked, mean over episodes | 98.77% | 98.77% |
| interpolated | 0.00% | 0.00% |
| in a long gap (masked) | 1.23% | 1.23% |
Frames with frame_valid = 1: 99.01% of all frames.
Notes for pi0.5 co-training
observation.stateandactionhold absolute poses. Chunk-relative actions (pose at t0+k expressed in the t0 frame) are meant to be computed in the training transform, because the horizon H is a training choice.loss_maskhas the same 16-d layout: multiply the per-dimension action loss by it (for a chunk, use the mask of each target frame).- Frame rate is the native 30 Hz for both datasets; nothing was resampled.
- No wrist cameras: use an image mask (False) for the two wrist slots.
- Quaternion y axis in world: median y·z_world per hand is -0.41 (left) and -0.20 (right), so thumb→index is mostly horizontal. On the robot, the tool y axis is close to vertical (left −0.93, right +0.90 against base z). Which robot tool axis the fingers close along is not verified (the URDF has no finger joints). If it turns out to be x, rotate the human frame by a fixed 90° about z in the training transform; the raw keypoints are included so the frame can be rebuilt differently.
pi0.5 training recipe (ego + robot co-training, and the robot-only baseline)
Values marked [set] were fixed by the project lead; [plan] come from the agreed co-training plan;
[openpi] are openpi defaults / its own pi0.5 fine-tuning configs (openpi main, checked 2026-09-26) and may be tuned.
Run the robot-only baseline with the same settings (only the ego share changes).
| item | value | source |
|---|---|---|
| init weights | gs://openpi-assets/checkpoints/pi05_base/params (full pi0.5 base) |
[plan] |
| batch size | 32 | [set] |
| batch composition | ego : robot = 1 : 1 in every batch (16 + 16); robot-only baseline = 32 robot | [set] |
| learning rate | constant after warmup: CosineDecaySchedule(warmup_steps=1_000, peak_lr=5e-5, decay_steps=30_000, decay_lr=5e-5) (openpi has no constant schedule class; pi05_libero uses this peak = end pattern) |
[set] constant; 5e-5 [openpi pi05_libero]; warmup 1k [openpi default] |
| optimizer | AdamW b1 0.9, b2 0.95, eps 1e-8, weight decay 1e-10, grad-norm clip 1.0 | [openpi] |
| EMA | 0.999 | [openpi pi05_libero] |
| train steps | 30,000 | [openpi default] |
| model | Pi0Config(pi05=True, action_horizon=50, discrete_state_input=False), action_dim 32 (14 used, rest zero-padded), max_token_len 200 |
[openpi] horizon/dims; no state [plan] |
| action | 14-d chunk-relative per arm: translation and rotation (rotvec) of the pose at t0+k expressed in the EE frame at t0, gripper absolute; computed from the absolute 16-d action column |
[plan] |
| state input | none for both embodiments | [plan] |
| normalisation | 1% / 99% quantiles to [-1, 1], per embodiment; ego stats exclude masked dims / frames; de-normalise with robot stats at inference | [plan] |
| loss mask | per-dim flow loss × loss_mask of each target frame (ego gripper dims and untracked hands = 0); drop windows touching frame_valid = 0 |
[plan] |
| discrete (FAST) loss | none for ego samples (openpi pi0.5 fine-tuning trains the flow loss only; if a FAST/discrete auxiliary is added, turn it off for ego) | [plan] |
| cameras | head → base image; left/right wrist → wrist slots; ego wrist slots image_mask = False | [plan] |
| wrist dropout (robot) | 15% per sample, drop one of the two wrist cameras at random; 0 for the last 10–20% of steps; all on at inference; same in baseline | [plan] |
| image augmentation | random crop + colour jitter (both embodiments) | [plan] |
| prompt | task column |
[plan] |
| seed | 42 | [openpi default] |
Conversion script: conversion/c10_convert.py in this repo.
- Downloads last month
- 887