angkul07/mf-pi05-teleop-isr

mf_pi05_teleop_isr — π₀.₅ on 100% ISR-standardized bimanual-handover teleop

Fine-tune of pi05_base (openpi flow-matching π₀.₅) on the ISR-resampled AgileX Piper bimanual handover dataset Kavin60606/bimanual-handover-isr-std (50 episodes / 78,379 frames / 20 Hz, task right_pick_handover_left_place). 100% teleop — no retargeted ego data in this run.

Recipe

• Config mf_pi05_teleop_isr (branch mf-piper-ea of angkul07/openpi): LoRA gemma_2b_lora trunk + full-rank action expert, action_dim=32 (padded from 14), action_horizon=50, max_token_len=200, augmentation on, EMA off, quantile norm.
• State/action: float32[14] = [right_j1-6, left_j1-6, right_gripper, left_gripper] — joints-major, grippers trailing — joints in DEGREES, grippers [0,1] (~1 open). Cameras top / left-arm / right-arm, 480x640 h264. The video is stored ~90° rotated, as captured; training and serving must agree on this geometry.
• 11,100 steps @ batch 64 (10.1 epochs over the 45 training episodes), cosine 3.5e-5 → 3.5e-6, warmup 1,000 steps. 2x H100 SXM data-parallel, 1.4–1.5 s/it, ~4.5 h wall-clock.
• Holdout: episodes 45–49 (contiguous tail, ~10%) never entered training — reserved for checkpoint-ladder scoring.
• Training pairs are RAW. The source dataset README suggests a t+2 control-delay realignment; measured on the ISR-resampled grid the effective lag is ~1 variable frame with sub-degree residual, and a t+2 shift fits worse than no shift, so none was applied.

Checkpoints

Each <step>/ holds params/ and assets/ (quantile norm stats for asset_id mf_teleop_isr — required at inference). train_state (optimizer state) is not included: these checkpoints are for inference/eval, not training resumption.

Ladder: 2000 / 4000 / 6000 / 8000 / 10000 / 11099 (final). Final training metrics (step 11099): flow_loss 0.0044, chunk_first 0.0014, chunk_last 0.0068, grad_norm 0.043.

Evaluation status

No policy evaluation has been run. Training loss alone supports no capability claim. The intended next step is scoring this ladder against holdout episodes 45–49.

W&B: https://wandb.ai/kavinrajkr60-dsfsd/mf-piper/runs/az78ujr1