PPO-run1
Preserved original Spatial PPO checkpoint at update 80: 63/100 success (63%) on paired LIBERO Spatial trials. This was the best measured checkpoint in the old extended PPO run. The selection score uses 100 trials and is not a 500-episode benchmark result.
update_080.pt is the original byte-for-byte PyTorch training checkpoint, including actor, value head, optimizer and per-rank RNG states. It requires the frozen VLM and saved processors from Riasok/few-shot-run-1; it is not a standalone Transformers checkpoint.
Load for inference/evaluation
Requires the MolmoAct2 training-enabled LeRobot environment used for this project. Download this repository and Riasok/few-shot-run-1 locally. Add this repository to Python's path, then:
import torch
from runtime.flow_ppo_pilot_v1 import load_sft_flow_adapter
adapter, bridge = load_sft_flow_adapter(sft_directory, noise_std=0.05, flow_steps=10)
saved = torch.load(ppo_directory + '/update_080.pt', map_location='cpu', weights_only=True)
adapter.actor.load_state_dict(saved['actor'], strict=True)
adapter.value_head.load_state_dict(saved['value_head'], strict=True)
adapter.to('cuda').eval()
Use the saved processor bridge for normalization and action unnormalization. Deterministic evaluation uses continuous flow inference, 10 flow steps and action chunks of 10, 256px agent/wrist images, 50 settling steps and horizon 280. See the included training driver for environment and evaluation details; its original absolute training paths need replacing before reuse.
eval_080.json, provenance.json, and training_config.json record the original evaluation and training configuration. export_manifest.json records file checksums and the HF source SFT dependency.