PPO-run2

Latest saved checkpoint of the second Spatial PPO experiment: update 107. The run was intentionally stopped on 2026-11-04 when the user switched to a fully frozen-VLM SFT -> PPO pipeline. It did not finish the originally planned 150 updates.

The starting SFT is Riasok/few-shot-run-2: 10 demonstrations/task (400 total), SFT step 5000, 402/500 (80.4%) Spatial success. That SFT trained VLM LoRA plus the action expert. This PPO run froze the adapted VLM and trained the action expert and critic.

Accuracy

The latest completed evaluation was update 100: 90/100 (90%). The best evaluated checkpoint was update 90: 93/100 (93%). The source SFT scored 81/100 on the same paired 100 trials. Update 107 itself has not been evaluated, so neither score should be attributed to these update-107 weights. See evaluation_summary.json and the original per-update evaluation JSON files. These are 100-trial results, not a final 500-trial PPO benchmark.

Contents and loading

update_107.pt is the complete original training checkpoint: actor, value head, optimizer, and per-rank RNG. It requires the frozen VLM and exact saved processors from Riasok/few-shot-run-2. Requires the MolmoAct2 training-enabled LeRobot implementation. Download both repositories locally and add this repository to Python's path:

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_107.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. The included training driver records the exact environment/evaluation implementation; original local paths need replacing before reuse. export_manifest.json records file integrity. checkpoint_metadata.json records verified checkpoint contents.