MolmoBot-Pi0 — DROID-toys, absolute joint, full fine-tune
Full (action-only) fine-tune of MolmoBot-Pi0-DROID
on shrg7/droid-toys, with
absolute joint-position action targets.
Checkpoint
- Step 5000 (end of schedule),
val_action_loss = 0.0079 - Flat
model.safetensors— loads directly into a plainPI0Pytorch assets/holds the norm stats;metadata.ptthe run metadata- Optimizer state is not included (inference/eval only, no resume)
Training
| Config | molmobot_pi0_lerobot_droid_absjoint |
| Dataset | shrg7/droid-toys (LeRobot v3.0) — 31 episodes / 6,952 frames @15fps |
| Action repr | joint_absolute — targets are future absolute joint positions q[t+1], 8-dim (7 arm + gripper) |
| Action horizon | 16 |
| Cameras | exterior_1_left (exo) + wrist_left, 224×224 resize-with-pad |
| Trainable | everything except vision_tower (frozen via freeze_filter) |
| Batch | 64 × 2 grad-accum = 128 effective, 4× NVIDIA L40 |
| LR | 5e-5 constant (200-step warmup; peak_lr == decay_lr, so the cosine is flat) |
| Steps | 5,000 (~50 epochs), 11h25m |
Validation curve
| step | val_action_loss |
|---|---|
| 3500 | 0.0096 |
| 4000 | 0.0086 |
| 4500 | 0.0085 |
| 5000 | 0.0079 |
Val loss was still descending at the end of the schedule — this checkpoint is not converged, and a longer run should improve on it.
Caveat
droid-toys is a small dataset (31 episodes). Treat this as a baseline for
comparison against the LoRA counterpart
(MolmoBot-Pi0-DROID-toys-absjoint-lora-r64),
which was trained on the same data with the same config, not as a
general-purpose policy.