PACT-WAM AR2 action-conditioned predictor
Exact step19690 checkpoint used by the PACT-WAM guidance experiments. Original precision and serialization are retained. WAM and JEPA encoder weights are not included. This is a custom predictor, not a standalone Transformers AutoModel.
Each predictor call consumes a current latent [B,576,1024], four low-level actions [B,4,7], robot state [B,8], and a phase index. Two autoregressive calls predict the two future horizons. Apply the supplied training-only action/state normalization and the camera/action contracts in the PACT-WAM source.
Implementation and evaluation: https://github.com/cpx196/PACT-WAM . Read the experiment reports for limitations: initial positive results did not extend to a broad success-rate improvement in the subsequent seven-task seed7 study. Optional task weights use base-task training demonstrations, not perturbed evaluation trajectories.
Files are indexed with SHA256 in artifact_manifest.json. Code entry point: code/fastwam_ar2_eval.py. Preserve the predictor checkpoint hash when using existing task calibration tables.