AgiBot G2 — pi05 H80 (action_horizon=80), step 10000
Fine-tuned π0.5 (pi05) policy from openpi (PyTorch training path) on AgiBot G2 right-arm teleop data for a three-object sorting task.
Files
model.safetensors— model weights (bfloat16, PaliGemma 2B VLM + 300M action expert)metadata.pt— full openpiTrainConfigsnapshot + global_step (torch.load,weights_only=False)assets/agibot_g2/g2_0611_lerobot_openpi/norm_stats.json— normalization stats (quantile norm), recomputed for the H80 episode-boundary sampling (sha2568bf450d0…)
Model / training config
| base checkpoint | pi05_base_pytorch |
| action_dim / action_horizon | 32 (8 used: right arm 7 joints + gripper) / 80 |
| state input | discrete (discrete_state_input=True), state dims 8–15 |
| images | head_color, hand_left, hand_right (224×224) |
| actions | delta joint actions (gripper absolute), quantile-normalized |
| batch / steps | 30 (6× RTX 5090) / 10000 of 30000 (cosine LR, peak 2e-5, warmup 1k) |
| vision tower | frozen (OPENPI_FREEZE_VISION_TOWER=1) |
| checkpoint step | 10000 — best validation checkpoint (val_loss 0.0135; val rises after ~5k due to small 229-episode train set) |
Data
Dataset agibot_g2/g2_0611_lerobot_openpi (LeRobot v3 format), 241 episodes / 111,345 frames @30fps, split 229 train / 12 val episodes.
Chunk sampling uses episode-level merged non-mistake spans (valid_interval_chunk_boundary="episode"): contiguous non-mistake intervals are merged, giving 60,888 train / 3,248 val H80 chunk starts (vs only 1,314/77 with the old per-interval boundary — the bug this run fixes).
Prompt
Trained with a single fixed prompt on all samples (prompt_from_interval=False); use exactly this prompt at inference:
抓取红色零件袋放在左边蓝色筐最远处的格子里，抓取黑色方块放在左边蓝色筐中间的格子里，抓取白色方块放在左边蓝色筐最近处的格子里
The model was never trained on per-sub-task prompts — do not expect sub-task-level language control.
Validation loss over training
| step | 5000 | 10000 | 15000 | 20000 | 25000 | 30000 |
|---|---|---|---|---|---|---|
| val_loss | 0.0120 | 0.0135 | 0.0155 | 0.0176 | 0.0188 | 0.0190 |
Runtime note
Although the policy predicts 80 actions per chunk, do not execute all 80 open-loop on a real robot — replan after a 10–30 step prefix.
Training run: g2_0611_pi05_rightarm_nonmistake_h80_episode_6x5090_bs30_20260706_022213, config pi05_agibot_g2_0611_h80, openpi fork branch fix/h80-episode-chunks (+ local fix: repack no longer requires a per-sample prompt key when prompt_from_interval=False).