Robot perception research artifacts

This is an exploratory experiment archive for an RSS research project. It is not a claim of acceptance, a finalized paper, or an official VER reproduction.

Code and protocol: https://github.com/YananZHOU5555/rss-robot-perception (private).

Weights: https://huggingface.co/B111ue/rss-robot-perception-checkpoints

Evidence: https://huggingface.co/datasets/B111ue/rss-robot-perception-evidence

This snapshot contains 25 completed legacy twenty-task ChunkBC readouts at the fixed 20,000th update, and 0 BC-ViLT policies at the fixed final epoch 30. BC-ViLT four-task calibration and ten-task comparison are separate cohorts. Selection does not use test success.

Each bundle has model.safetensors, config.json, and export.json. Frozen upstream vision encoders are not included. BC-ViLT additionally includes fixed BERT task_embeddings.safetensors, checkpoint_metadata.json with task/data identities, and its own restore_checkpoint.py. The native model identifiers, preprocessing, split and source hashes remain in the bundle metadata. Optimizer and RNG state are retained in the local experiment archive, but excluded here.

Restore a legacy ChunkBC bundle:

python restore_checkpoint.py checkpoints/RUN_NAME --output restored.pt --campaign-root /path/to/campaign

Restore a BC-ViLT bundle for vilt_eval.py:

python checkpoints/RUN_NAME/restore_checkpoint.py checkpoints/RUN_NAME --output restored.pt --campaign-root /path/to/campaign

Restoration verifies tensor and file hashes. The reconstructed checkpoint has identical learned tensors but a different file hash from the original training checkpoint.

No project-wide redistribution license is assigned in this archive; upstream assets retain their original terms.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading