eugene123tw/physical-ai-studio-push-t

Pi0.5 PushT Inference

Evaluate a trained Pi0.5 checkpoint on the PushT environment.

Prerequisites

Install the physicalai library with gym dependencies:

uv sync

A CUDA-capable GPU is required.

Usage

python pi05_pusht_eval.py --ckpt <path-to-checkpoint> [--n-episodes N] [--output-dir DIR]

Arguments

Argument	Required	Default	Description
--ckpt	Yes	—	Path to the .ckpt model checkpoint
--n-episodes	No	5	Number of evaluation episodes
--output-dir	No	./tmp_scripts/pusht_eval_videos	Directory for recorded episode videos

Example

python pi05_pusht_eval.py \
    --ckpt ./pc_success=90.00.ckpt \
    --n-episodes 10 \
    --output-dir ./eval_videos

Output

The script prints aggregated metrics (success rate, reward, episode length, FPS) and a per-episode breakdown, then saves video recordings of each episode to the output directory.