brandonyang/diffusion_policy_robocasa_activations_latest_chkpt

Diffusion Policy — RoboCasa Activations (latest checkpoint)

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 "Using with diffusion_policy" below.

What's in the bundle

7 RoboCasa tasks (~30 episodes each, mix of successful and failed rollouts):

• CloseFridge
• CoffeeSetupMug
• OpenDrawer
• OpenStandMixerHead
• PickPlaceCounterToCabinet
• PickPlaceCounterToStove
• TurnOnElectricKettle

Plus success_rates.json summarizing per-task rollout outcomes.

Note: CloseBlenderLid was intentionally excluded from this upload.

Total size: ~327 GB.

Repo layout

<repo-root>/
  success_rates.json
  <task>/                                  # one of the 7 tasks above
    episode_NNN_env_NNN/
      metadata.json                        # episode_success, total_reward, steps_to_success, ...
      rewards.npz                          # per_step_reward, cumulative_reward, success_at_step
      step_NNNN/                           # one folder per env step (every n_action_steps)
        metadata.json                      # task_name, episode_id, step, inference_step, prompt, ...
        denoising.npz                      # all_x_t (D, H, A), all_v_t (D, H, A)              fp32
        adarms_cond.npz                    # all_adarms_cond (D, T_cond, C)                    fp32
        suffix_residual.npz                # all_suffix_residual (D, L, H, C)                  fp32

Schema id: dp_v1 (stamped in each step's metadata.json under collection_version). For the captured model: D=100 denoising steps, L=12 decoder layers, H=10 action tokens, C=512 hidden dim, A = action dim.

Key files

file	array key	shape	purpose
step_*/suffix_residual.npz	all_suffix_residual	(D, L, H, C)	post-layer residual stream — primary signal for conceptor / SAE steering
step_*/denoising.npz	all_x_t, all_v_t	(D, H, A)	denoiser input/output — useful for action-space ablations
step_*/adarms_cond.npz	all_adarms_cond	(D, T_cond, C)	encoder cond-token sequence — useful for prompt/cond ablations
episode_*/rewards.npz	per_step_reward, cumulative_reward, success_at_step	(T,)	per-env-step reward history
episode_*/metadata.json	—	—	episode_success label (used to split success/failure for contrastive conceptors)

Downloading

Option 1 — huggingface_hub (recommended)

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="brandonyang/diffusion_policy_robocasa_activations_latest_chkpt",
    repo_type="dataset",
    local_dir="diffusion_policy/activations/latest",   # matches the default layout
    max_workers=8,
)

Option 2 — hf CLI

hf download brandonyang/diffusion_policy_robocasa_activations_latest_chkpt \
    --repo-type dataset \
    --local-dir diffusion_policy/activations/latest

Option 3 — selective download (single task)

If you don't need all 327 GB, grab one task at a time:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="brandonyang/diffusion_policy_robocasa_activations_latest_chkpt",
    repo_type="dataset",
    local_dir="diffusion_policy/activations/latest",
    allow_patterns=["OpenDrawer/**", "success_rates.json"],
)

Using with diffusion_policy

The activation tree is the input format expected by the steering pipelines in the diffusion_policy repo. After downloading into diffusion_policy/activations/latest/:

# Build contrastive conceptors over the success/failure split
python experiments/robocasa_steering/build_conceptors.py \
    --activations-dir diffusion_policy/activations/latest

# Or train an SAE on the residual stream
python experiments/sae/src/train_sae.py \
    --activations-dir diffusion_policy/activations/latest

Both consumers read <task>/episode_*/metadata.json (for the success label) and <task>/episode_*/step_*/suffix_residual.npz (key all_suffix_residual, shape (D, L, H, C)). They mean-pool over the H action tokens and stack across episodes.

If you want to point at a different download location, pass --activations-dir /your/path (or set DP_ACTIVATIONS_DIR=/your/path).

Provenance

• Model: DiffusionTransformerHybridImagePolicy (12-layer transformer decoder, 100-step DDPM).
• Checkpoint: latest.ckpt (final training step).
• Collector: collect_activations_robocasa.py (dp_v1 schema).
• Hardware: EGL-rendered MuJoCo, single-GPU collection.

Citation

If you use this data, please cite the underlying diffusion_policy and RoboCasa works.