TeleArms YAM cubes v5: realistic domain randomization
The TeleArms teleoperated demonstrations (bimanual YAM cube manipulation, MuJoCo): the
demos of mliu48/telearms-yam-cube-v4-dr
and the in-browser local-play demos recorded since 2026-08-06, and their synthetic colour
and letter variants. Every episode is re-rendered at the real cameras
under domain randomization dressed like the real setup. LeRobot v3.0, 30 Hz.
| Episodes | Frames | |
|---|---|---|
Human demos, realistic (<uuid>-d1) |
236,371 | 290,204,101 |
Human demos, viewpoint (<uuid>-d2) |
236,371 | 290,204,101 |
Human demos, appearance (<uuid>-d3) |
236,371 | 290,204,101 |
Synthetic variants, realistic (<variant>-d4) |
156,067 | 179,308,077 |
| Total | 865,180 | 1,049,920,380 |
Each of the 236,371 human demos appears three times, with the same states and actions
under three different renderings; each synthetic variant appears once.
meta/episode_sources.json lists which episodes are human and which synthetic.
Cameras
Each camera renders at 640×360 (16:9) with its real vertical field of view (RealSense intrinsics) and is turned by its real principal-point offset, then is letterboxed to 224×224 with antialiasing (the full frame 224 wide, centred black bars, no crop), as the real robot does with its frames. The top camera sits on its fixed mount 0.95 m above the table, 15.6° off vertical, centred between the arms, lens above the arm bases (the measured 12 Oct pose, raised from 0.72 m so that no cube leaves the view).
| Key | Real camera | Vertical field of view |
|---|---|---|
observation.images.camera_front |
RealSense D435IF colour | 43.1° |
observation.images.camera_left, camera_right |
RealSense D405 (wrists) | 58° |
meta/info.json records camera_resolution: [640, 360]. Feed real frames as 224×224
letterboxes made with an antialiased (area) resize of the full 1280×720 frame
(sim2real/real_cameras.letterbox); the policy's own resize then never runs. For pi0.5, the keys map to base_0_rgb, left_wrist_0_rgb and
right_wrist_0_rgb.
Domain randomization
Each episode draws its randomization once and keeps it for every frame. Physics and actions are untouched. No per-frame sensor noise is baked in: add it as training-time augmentation.
The top camera's fixed mount is set by hand, so every profile tilts it up to ±5° up or down and pans it up to ±3° around that pose.
Scene, every episode. The table top is one of seven styles, each in its own colour range, mostly wood (80 %) since policies are tested on a wooden table: light oak like the real table (49 %), maple (13 %), walnut (9 %), MDF (9 %), dark laminate (8 %), white laminate (8 %), brushed steel (4 %). Under the arms, table top and a dark aluminium extrusion as on the real cell (render only); behind the table a black curtain, below it a dark floor.
| Profile | Lights | Cameras | Colours and sensor |
|---|---|---|---|
Realistic (-d1, -d4) |
±20 cm, ±12°, brightness ×0.6–1.2 | top ±2 cm, ±2°, tilt ±5°, pan ±3°; wrists ±0.5 cm, ±2°; field of view ±2° | — |
Viewpoint (-d2) |
as realistic | top up to ±15 cm per axis, re-aimed at the table, ±8°, tilt ±5°, pan ±3°, field of view ±8°; wrists ±2 cm, ±8°, ±5° | mild exposure, white balance, gamma, contrast, saturation |
Appearance (-d3) |
±50 cm, ±35°, brightness ×0.3–1.7, tint ±25 % | top ±4 cm, ±4°, tilt ±5°, pan ±3°; wrists ±0.8 cm, ±4° | cube and patch colours jittered (still recognizable); strong exposure, white balance, gamma, contrast, saturation, blur |
Synthetic variants
A variant replays an original demo with its cubes, patches or letter faces swapped
one-to-one for other colours (red, blue, green, yellow, purple, orange) or letters,
and its prompt rewritten to match. The motion is unchanged, so the variant succeeds
at its new task by construction. Symmetric letters (H I N O S X Z) swap only among
themselves; word tasks swap to real words; alphabetical rows stay in order.
task_name is <original task> (relabeled); each distinct prompt is one synthetic task.
Prompts
meta/task_prompt_variants.json (format 2) holds 50 paraphrases per original task
and, for each synthetic task, its colour and letter mapping. Synthetic tasks reuse
their original task's paraphrases with the mapping applied. The original instruction
stays in every pool.
State and action
State and action are 14-dimensional: 6 arm joints plus 1 gripper per arm. Actions
are absolute joint targets. meta/stats.json holds normalization statistics in the
50-step relative space the training recipe uses (arm deltas to the chunk anchor,
absolute grippers); do not renormalize relative actions with absolute statistics.
- Downloads last month
- 84