wjwjwjwj/robotwin-adjust_bottle-clean50-randomized50-gr00t

RoboTwin 2.0 adjust_bottle: clean 50 + randomized 50 (GR00T N1.7 format)

Training set for GR00T N1.7 in LeRobot v2.1 format: 100 episodes, 14,283 frames, 25 fps metadata. This is the data mix that worked best in our augmentation comparison on RoboTwin 2.0 adjust_bottle.

GR00T N1.7 training data (same recipe)	Easy	Hard
clean 50 (baseline)	73	52
clean 50 + randomized 50 (this dataset)	96	87
clean 50 + Cosmos3 video augmentation 49	97	48
clean 50 + scripted recovery demos 50	83	29

Success rate (%) over 100 test episodes each (RoboTwin eval_policy.py, seed 0, unseen instructions). Hard = unseen wall/table textures, up to 10 distractor objects, random light color.

What is in it

• Episodes 0-49: the 50 official RoboTwin 2.0 demo_clean expert demonstrations of adjust_bottle (aloha-agilex).
• Episodes 50-99: the 50 official RoboTwin 2.0 demo_randomized expert demonstrations (random wall/table textures from the training texture set, up to 10 distractor objects, random light color, table height jitter).
• The demonstrations are RoboTwin's own data (MIT license). We converted them to GR00T format and combined them; no frames are generated or edited.

Format

• Cameras: cam_high (head), cam_left_wrist, cam_right_wrist, 320x240 RGB.
• observation.state / action: 14-dim absolute joint positions (left arm 6, left gripper 1, right arm 6, right gripper 1). Gripper 0 = closed, 1 = open.
• action[t] = qpos[t+1]; the last frame of each episode is dropped.
• Language: one of the 50 RoboTwin "seen" instruction templates, sampled per frame.
• meta/modality.json: arms as RELATIVE (6 dims each), grippers as ABSOLUTE (1 dim each).
• meta/stats.json, meta/relative_stats.json: computed over all 100 episodes.
• Correct-arm balance: left 51 / right 49 (clean 50 alone is 31 / 19).

Training recipe used for the numbers above

GR00T N1.7-3B (nvidia/GR00T-N1.7-3B), NEW_EMBODIMENT, backbone frozen (projector + action head trained), 10,000 steps, global batch 64, checkpoints every 2,500 steps. The checkpoint was chosen on a separate validation split (seed 1, Hard, 50 episodes) and then tested on seed 0. The selected checkpoint for this dataset was step 7,500.

Limitations

One task, one training run per data mix. The 95% interval of a 100-episode success rate is about ±8-10 %p. The gain mixes two effects: scene randomization and the correct-arm balance (31/19 → 51/49).

Source

• RoboTwin 2.0: https://github.com/RoboTwin-Platform/RoboTwin (data: TianxingChen/RoboTwin2.0)
• Reproduction context: "Seeing Realism from Simulation" (ICML 2026, arXiv 2605.02757)
• Maintainer: Wonjin Shin (Nota AI, XPU team)