Per-step, per-episode activation traces collected from a DiffusionTransformerHybridImagePolicy (the diffusion_policy library) rolled out on RoboCasa benchmark tasks. Captured with collect_activations_robocasa.py at the latest training checkpoint.
These traces are the input expected by the conceptor / SAE steering pipelines under diffusion_policy/experiments/robocasa_steering/ and diffusion_policy/experiments/sae/ — see… See the full description on the dataset page:
https://huggingface.co/datasets/brandonyang/diffusion_policy_robocasa_activations_latest_chkpt.
