Public PyTorch checkpoints trained on the ForceBenchmark planning datasets with Panda qpos(9) + qvel(9) 18D state, 7D pd_ee_delta_pose actions, and 6D incoming_joint wrist wrench.

Protocol: seed 1, batch size 256, observation horizon 2, action horizon 8, prediction horizon 16, 100,000 training iterations, train-only (no evaluation). Each .pt stores both agent and ema_agent state dictionaries.

Included: all five Lift checkpoints and all five Stamp checkpoints (Exp2/4/5/6/7).

Force settings: Lift uses an 8-frame wrench window and normalized clip 5; Stamp uses a 4-frame wrench window and normalized clip 6. See metadata/training_manifest.txt and metadata/force_stats/ for details. Verify files with SHA256SUMS.