visual_robust_libero — Experiment 08: static embodiment supervision

Data for the question "can supervising a policy on static embodiment information (where the gripper is, which arm it is) improve cross-embodiment task transfer?", built on LIBERO with robosuite 1.4.0 / MuJoCo 3.3.7.

24 embodiments = 6 robots x 4 grippers. Four of them carry action demonstrations; the other 20 appear only as static or demo-free synthetic data, which is what makes the transfer question askable.

Layout

action/                LeRobot, 8 datasets, 294 episodes, 47,031 frames
synthetic_replay/      LeRobot, 8 shards, 48 episodes, 8,602 frames x 24 embodiments
synthetic_randik/      LeRobot, 4 shards, 96 episodes, 23,040 frames
vqa/{b1,b2,c1,c2,d}/   LLaVA-style instruct.json (images not shipped -- see below)
metadata/              base positions, embodiment table, 13,991 motion pairs
scripts/               generators, auditor, VQA builder

The VQA image files are deliberately not uploaded. They are JPEG copies of frames that already sit inside the parquets, and b1/b2 shared one identical set while c1/c2 shared another -- 637 MB of pure duplication on top of the originals. Regenerate them from the parquets:

python scripts/exp08_build_vqa.py --root <this directory> --out vqa \
  --action-stride 4 --replay-stride 12 --randik-stride 2 --horizon 10

That reproduces all 116,800 examples (b1 11,864 / b2 11,864 / c1 23,520 / c2 47,040 / d 22,512). The shipped instruct.json files are the exact ones those settings were built with, so questions and answers need no regeneration.

Settings

Setting	Source	Question -> answer
A	action only	(vanilla, no VQA)
B1	real action frames	end-effector position, base-relative
B2	real action frames	end-effector pixel
C1	synthetic, 24 embodiments	end-effector position, base-relative
C2	synthetic, 24 embodiments	end-effector pixel + embodiment identity
D	synthetic image pairs	movement direction + distance

Embodiments

Robots: Panda, IIWA, UR5e, Jaco, Kinova3, Sawyer. Grippers: Panda, Rethink, Robotiq 2F-85, Robotiq 2F-140.

Action demonstrations exist only for:

Robot + gripper	Suite	Tasks
IIWA + Robotiq85	libero_spatial	0, 1
Panda + Panda	libero_object	0, 1
UR5e + Robotiq85	libero_goal	4, 8
Panda + Rethink	libero_10	2, 8

Every embodiment carries a category relative to those four — training, seen_robot_seen_gripper, seen_robot_unseen_gripper, unseen_robot_seen_gripper, unseen_robot_unseen_gripper — and every VQA example repeats it in meta.category, so a held-out transfer analysis is a filter rather than a re-derivation from tag strings.

Coordinates

Use observation.eef_base_rel, not the world position. libero_object mounts the arm at [-0.60, 0, 0.000] while the other three suites mount it at [-0.66, 0, 0.912], so world z differs by 0.91 m for a visually identical pose. A model asked to predict world z would have to identify the scene first, which is exactly the shortcut this experiment measures. Subtracting each arm's own base collapses the suites onto each other (median z spread 90 cm -> 2 cm) while preserving the genuine per-robot differences (at rest, eef-base z is 0.261 for Panda but 0.093 for Kinova3). Both frames are stored; metadata/base_positions.json holds the offsets.

eef_pixel is in stored-image coordinates. Frames are np.flipud of the raw render and robosuite's project_points_from_world_to_camera already returns top-down rows, so do not mirror the row again — the two flips cancel. A wrong convention still yields 100% "in frame" and plausible ranges; it only shows up when the marker is drawn on the image.

Two synthetic variants, and why both

synthetic_replay drives all 24 embodiments along one demonstration's end-effector path, so a frame shows the same pose on 24 different arms. That alignment is what makes "same motion, different robot" comparisons possible. Measured cross-embodiment displacement spread is a median of 1.3–2.7 cm (up to 16.8 cm on long-horizon shards), so each embodiment carries its own measured displacement rather than a shared label; cross_embodiment_spread_cm in metadata/motion_pairs/ lets the loose pairs be filtered out.

synthetic_randik samples end-effector targets directly, with no demonstration, solving each arm's IK for a pixel-and-depth target so every accepted pose is visible. It covers roughly 5x the volume (convex hull 1,502 L vs 278 L; bounding box 148x124x149 cm vs 48x78x130 cm).

They are complements, not alternatives. Replay motion directions are severely skewed — on libero_spatial_t0, down 470 / forward 435 versus left 32 / right 7, because a demonstration is mostly reach-down-and-push-forward. Random-IK steps are drawn isotropically and come out near-uniform (434–533 per axis). Setting D trained on replay alone would let a model score well by always answering "down". The shipped vqa/d/instruct.json mixes both.

Gotchas when regenerating

• Sawyer fails to compile ("mesh volume is too small: robot0_head_1") unless head* meshes get inertia="shell" at load time — see scripts/exp08_libero_compat.py.
• env.reset() frees and rebuilds the sim, leaving any cached sim, site id, or joint index dangling. It surfaces as 'MjSim' object has no attribute 'data'.
• Gripper state widths differ per gripper (Panda 2, Robotiq140 6, Robotiq85 8 joints), so read the pose from the end of observation.state ([-7:]), never at a fixed offset. observation.gripper_open is the normalised, cross-gripper comparable scalar.
• IK ignores collisions, so a solved target can put the gripper under the table. synthetic_randik rejects those with a mujoco.mj_ray visibility test from the camera to the grip site.

Verification

scripts/exp08_audit.py reports 0 failures / 0 warnings over all three families: every replay episode carries all 24 embodiments with decodable images, and end-effector pixels are in frame for 100% of frames everywhere.