agi
Collection
Robot policies, served over an API. agi.diy's own SO-101 checkpoints (ACT, SmolVLA, Cosmos3-Edge WFM), held-out numbers, 4-bit exports. api.agi.diy • 3 items • Updated
10k episodes · 15 fps · 2 cameras · 640×480 av1
so101-ttt-sim2real-cosmos3
10,000 simulated SO-101 tic-tac-toe pick-and-place episodes (1,351,748 frames, 25.0 h at 15 fps, 12.1 GB). Trajectories, tasks and timestamps come from MuJoCo unchanged; both cameras were re-rendered with nvidia/Cosmos3-Nano-Sim-Transfer to look like one real bench. Generated video, not recordings of a real arm. Same schema as cagataydev/so101-ttt-flywheel-sim.
Load
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("cagataydev/so101-ttt-sim2real-cosmos3", revision="v3.0")
Schema (LeRobot v3.0)
| key | shape | notes |
|---|---|---|
action, observation.state |
float32[6] | shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll in degrees; gripper 0-100 % |
observation.images.top, .wrist |
480x640 | AV1, 15 fps |
| task | 38 strings | "pick up the x piece and place it on the top right cell" |
meta/flywheel.jsonl |
10,000 rows | source_episode, seed, prompt, seg_remap, align, tone |
How it was made
- MuJoCo bench scene, scripted IK expert, geometry-only randomisation; successful episodes only.
- Depth, segmentation and edge controls per camera (
so101-ttt-sim2real-controls); raw sim RGB (so101-ttt-sim2real-sim-rgb); same episode ids. - Cosmos3-Nano-Sim-Transfer on 6 Hugging Face Jobs H200 shards, segmentation control, one prompt per episode.
- Assembled by the flywheel-sim LeRobot writer; action/state/timestamp rows byte-equal to the source.
Populations (seg_remap, prompts_used)
| population | episodes |
|---|---|
| prompts-v4, palette remap (white O rings) | 7,590 |
| prompts-v4, no remap (held O ring often cyan on the wrist) | 2,381 |
| prompts-v3 (arm black) 26; wrist-only remap 3 | 29 |
QA (100 episodes, details in QA.md)
Labels byte-equal 100/100. FID vs real, raw sim to cosmos: top 391 to 279, wrist 314 to 168. NN-LPIPS: top 0.430 to 0.461, wrist 0.693 to 0.575. Control-edge misplacement top 3 %, wrist 18 %. Colour semantics 97/100.
Results
Three post-tunes of cagataydev/agi-wfm-so101-ttt on this data (mix ratio 0.5 or 1.0; 1,500 or 2,500 iterations) did not beat the served policy in closed-loop sim play: pooled 3-seed boards 413/540 vs 411/540 (p=0.94). Clean-up 7/90 vs 3/90 (small n).
Known limits
- Wrist view: held O rings sometimes render as rope or tape; X pale lilac in about 10 % of episodes; rare hallucinated objects.
- Labels are sim expert trajectories; no real-arm success is implied.
- Top-camera NN-LPIPS got worse; only FID improved.
License
Apache-2.0. Generated with NVIDIA Cosmos3-Nano-Sim-Transfer (NVIDIA Open Model License). Cite: Cagatay Cali, so101-ttt-sim2real-cosmos3, 2026, https://huggingface.co/datasets/cagataydev/so101-ttt-sim2real-cosmos3
- Downloads last month
- 101