ACT-so101-pickplace-sim-150-10K-v1.0

Completed ACT policy trained on local dataset physicalai/so101_pickplace_sim_150 for 10,000 optimizer steps. Published 2026-11-13. Recovered from Runpod Global Volume /shared-volume/act_so101_150_10k; final checkpoint was selected on seed 1000 and evaluated independently on seed 2000.

This is a native LeRobot 0.6.1 policy package, including the original model weights, policy configuration, preprocessor/postprocessor definitions and normalization tensors. It predicts 6-dimensional joint actions from 6-dimensional state and top/wrist RGB cameras (3 x 240 x 320), with action chunk size 50. Camera names and normalization must match the saved configuration. Architecture: ACT with a ResNet18 backbone initialized from ImageNet weights, Transformer encoder/decoder and CVAE.

Loading

Install lerobot==0.6.1 in a compatible environment. Download a pinned snapshot with hf download caspar/ACT-so101-pickplace-sim-150-10K-v1.0 --revision v1.0 --local-dir ./policy, then load it with ACTPolicy.from_pretrained("./policy") from lerobot.policies.act.modeling_act. Use LeRobot's policy pre/postprocessors from that same directory for observation normalization and conversion of predicted actions to joint units. Loading the neural network alone does not apply those processors. Training/evaluation scripts and simulator are in physicalai_tutorial.

Evidence and limits

See training_summary.json and evaluation.json for original run settings and historical simulation results. Publication verified original checkpoint SHA256, processor assets, and finite tensors. No new training, policy inference or simulator rollout was run for this upload. Historical simulation success rates are not real-robot validation; keep them tied to the recorded seeds and success criterion. The original run uses the local dataset identifier physicalai/so101_pickplace_sim_150; its dataset revision and episode count are not established by the recovered checkpoint. Do not assume it matches the later caspar dataset revision. The 10K checkpoint is a separate run from the 2500-step comparison. No optimizer states, source demonstrations or credentials are included.