pi0-FAST DK1 LoRA Fine-tune v1
Fine-tuned pi0-FAST on egocentric manipulation data for the TRLC DK1 Follower robot arm.
Training Details
| Parameter | Value |
|---|---|
| Base model | pi0-FAST (PaliGemma 2B LoRA) |
| Fine-tune method | LoRA (gemma_2b_lora) |
| Steps | 1000 |
| Batch size | 32 |
| Action dim | 7 (6 arm joints + 1 gripper) |
| Action horizon | 10 |
| Max token length | 350 |
| delta joint actions | True |
Dataset
| Field | Value |
|---|---|
| HF Dataset | kavinrajkr60/dk1_egodex |
| Source | EgoDex egocentric manipulation episodes |
| Episodes | 1800 (enriched ego part1 + part2) |
| Annotations | L1-L5 language annotations |
| FPS | 10 |
| Robot | TRLC DK1 Follower (6-DOF + parallel jaw gripper) |
Robot: TRLC DK1 Follower
- 6-DOF revolute arm + 2 prismatic gripper joints
- End effector: tool0
- IK solver: PINK (Pinocchio, QP-based differential IK)
- Action space: 7-dim [joint1..joint6, gripper_open]
Loss Curve
- Step 0: loss=12.56
- Step 300-360: loss=3.21-3.38
- Step 1000: training complete (~55 min, single GPU)
Loss dropped from 12.56 to ~3.2 over 1000 steps.
W&B Run
https://wandb.ai/kavinrajkr60-dsfsd/roboinc/runs/22za2p1i
Known Issues
Token truncation: some samples exceed max_token_len=350 (up to 414 tokens). Increase in next run.