pi05_xarm_cube_points9_arrow_len0_lora
pi0.5 (openpi) LoRA fine-tune for the xArm cube task ("pick up the cube"), trained on
camera frames with the points9_arrow_len0 tactile overlay: 9 tactile-pad dots per
finger at FK-projected positions, force arrows at zero length — the force-information
ablation of points9_arrow.
Training
| |
|---|
| Base | pi05_droid (knowledge-insulated pi0.5, DROID pre-training) |
| Recipe | LoRA r16/a16 on Gemma-2B backbone + r32/a32 on 300M action expert, attn+ffn |
| Batch / LR | 8 / cosine 1e-4 -> 1e-5, 500 warmup, AdamW, grad-clip 1.0 |
| Data | EdwardoSunny/xarm_cube_points9_arrow_len0 (100 eps, 13103 frames) |
| Early stop | step 7900 (rolling-1k-window <0.5% rel. improvement, 2 consecutive checks) |
| Final train loss | 0.0435 (baseline no-overlay: 0.0437; full points9_arrow: 0.0425) |
Training-loss comparison across this task's variants (identical hyperparameters/protocol,
only the burned-in overlay differs) — see the collection for all four tasks.
Contents
params/ — merged base+LoRA weights (orbax), self-contained for inferenceassets/local/xarm_cube_points9_arrow_len0/norm_stats.json — state/action normalization- Optimizer state stripped (inference-ready). Config name:
pi05_xarm_cube_points9_arrow_len0_lora
Deployment note
At inference, render the live camera frames with the SAME overlay
(SensorDrawer mode points9_arrow, arrow_length_scale=0, thickness 8, dot size 22)
or the policy will be out-of-distribution. See tactile-data-collection/scripts/render_arrowlen0.py.
