polaris_pi05_steerable_policy

Fine-tuned π0.5 (pi05) DROID steerable policy for the Polaris real2sim stack. Trained with TAMP-derived steering commands so the policy can be conditioned on sub-goal / steering instructions at inference time.

• Base config: pi05_droid (openpi)
• Recipe: pi05_droid_steerable_lerobot_finetune_polaris
• Checkpoint step: 19999 (final)
• Contents: inference-ready only — params/ + assets/droid/norm_stats.json. The optimizer train_state/ is not included, so this checkpoint serves but cannot resume training.

Usage

huggingface-cli download Makolon0321/polaris_pi05_steerable_policy --local-dir /path/to/ckpt
export POLARIS_PI05_CKPT=/path/to/ckpt

The repo root acts as the openpi checkpoint dir; create_trained_policy (pi05_droid config) reads params/ and assets/ from it. See docs/steerable_sft.md in the polaris_real2sim repo for the steering-command protocol.