fbarbosa1/pi05-driving-grpo-5k-checkpoint

pi0.5 Driving — Flow-GRPO post-trained (grpo-5k-v2/5000)

Flow-GRPO post-train of markmusic/pi05-driving-bc-v2-checkpoint on NVIDIA PhysicalAI-AV (Alpamayo) driving clips.

Held-out validation (10 clips, vs the BC-v2 base)

Model	ADE (m) ↓	val_loss (flow) ↓
BC-v2 (base)	3.56	0.237
GRPO-5k (this model)	2.81 (−21%)	0.190 (−20%)

GRPO improves both displacement and the held-out flow-matching loss, and (unlike the GT-imitation DPO variant) it does not over-steer on turns.

Method

Composite reward = PDMS driving quality (progress / comfort / drivable-area / time-to-collision)

• language-command consistency + a clipped GT-proximity guardrail. Group-relative advantages (group mean = baseline), PPO-clipped importance ratio, KL-anchored to the frozen BC reference. Flow-policy log-prob via a flow-matching-MSE surrogate. 5000 steps, peak LR 1e-6, KL coef 1.0.

Load (openpi)

from huggingface_hub import snapshot_download
import openpi.training.config as config
import openpi.policies.policy_config as policy_config

ckpt = snapshot_download("fbarbosa1/pi05-driving-grpo-5k-checkpoint")
cfg = config.get_config("pi05_driving")
policy = policy_config.create_trained_policy(cfg, ckpt)

Contains params/ (weights) and assets/ (norm stats). Optimizer state (train_state/) omitted (inference-ready).