SemWAM RoboDojo v1

Final model checkpoint for RoboDojo, with the saved inference contract and matching numerical normalization statistics.

Files

  • checkpoints/weights/step_100170.pt: v2 model weights, including the frozen semantic compressor.
  • inference_config.json and config.yaml: saved inference configuration.
  • dataset_stats.json: matching numerical action/state statistics.
  • checkpoint_manifest.json: file checksums and required frozen asset hashes.
  • model_card_summary.json and trainer_state.reference.json: inference reference metadata.

Inference

Use the SemWAM runtime to restore the saved v2 contract before constructing the model and loading weights. Prepare external frozen assets matching the saved hashes; the semantic compressor is embedded.

Inputs are three current RGB-D camera observations, a task instruction, and a 14-dimensional joint state. Supply native uint16 depth in millimetres with matching camera intrinsics. Output actions follow left-arm six joints / left gripper / right-arm six joints / right gripper order. Preserve the supplied normalization and preprocessing.

This package contains no optimizer, per-rank RNG, full training state, training data provenance, or private paths. Publication does not establish closed-loop benchmark performance.

Integrity

  • Weight SHA256: eae6522f94d0c81d66883fb8799760c931b6b4ee7d3f9e1b6e49d44eb4fab76e
  • Statistics SHA256: 06dad89bc0b437dd664c5de095461b8237e9b68db413410f70b5b7b2e2299af3
Downloads last month
27
Video Preview
loading