podolinsky/pi0.5-libero-10-features

pi0.5 LIBERO-10 prefix features — normal + occluded

Aligned rollouts of the stock openpi pi05_libero policy on the LIBERO libero_10 suite, collected in two matched scene variants:

• normal — the original LIBERO libero_10 tasks / initial states.
• occluded — the filename- and initial-state-matched libero_10_occluded suite from LIBERO-Occ, which adds scene-induced occlusion to the same tasks so the effect of occlusion can be read off pairwise.

10 tasks × 2 variants × 25 initial states = 500 episodes (25 normal + 25 occluded per task). Every π0.5 inference, executed action, and video frame in an episode is joinable by array row.

Benchmarks

• LIBERO — Liu et al., Benchmarking Knowledge Transfer for Lifelong Robot Learning, arXiv:2306.03310
• LIBERO-Occ — Li et al., Evaluating and Improving Vision-Language-Action Models under Scene-Induced Occlusion via Viewpoint Imagination, arXiv:2606.10862
• π0.5 — Physical Intelligence, a Vision-Language-Action Model with Open-World Generalization, arXiv:2504.16054

Which π0.5 features

π0.5 runs a prefix forward pass through its PaliGemma backbone (SigLIP vision + Gemma-2B LM) over the image and language tokens; the resulting last-layer hidden states fill the KV cache that then conditions the flow-matching action expert. openpi normally discards those hidden states — here they are captured, once per policy inference (i.e. every replan_steps = 5 control steps), raw and per-token (no pooling), as float16:

key	shape	tokens
base_image	(256, 2048)	base RGB camera, 224×224 → 16×16 SigLIP patches
wrist_image	(256, 2048)	left wrist RGB camera, same
language	(L, 2048)	instruction tokens (padded to 200; see language_mask)
language_mask	(L,) bool	real vs padding for language

2048 is the Gemma-2B hidden width. The always-zero right-wrist camera slot is dropped. These are the frozen-backbone representation before the action expert — the input to any perception / failure probe.

Layout

<scene_variant>/<NN>_<task_stem>/ep<NNN>/
    rollout.json    metadata + per-policy / per-control clock records
    rollout.npz     the arrays below
    rollout.mp4     agentview video, 20 fps, one frame per control step
    wrist.mp4       wrist video, same timing
    labels.json     Gemini video description + failure localization (see Labels)
    labels.npz      the same phrases / failure fields as aligned arrays
    example.md      human-readable render of labels.json
manifest.csv        one row per episode
gemini_prompts.md   verbatim Gemini prompt templates (+ gemini_prompts.json)

rollout.{json,npz,mp4} and wrist.mp4 are written once at collection and are never touched again; the label files are a cheap additive diff (a few KB per episode). All 500 / 500 episodes are labeled.

rollout.npz

Features: base_image, wrist_image, language, language_mask — each stacked over the n_policy inferences.

Actions: predicted_action_chunks (n_policy, 10, 7), predicted_chunk_len, executed_actions (n_control, 7).

Clocks, per control step (n_control,): control_step, sim_step (includes the settle/wait steps, which are not in the video), policy_step (which inference produced this step), chunk_index (position within that predicted chunk), video_frame_id (== control_step).

Scalars: success, replan_steps, n_policy, n_control, img_tokens, hidden, control_hz (20).

Alignment contract

executed_actions[t] == predicted_action_chunks[policy_step[t], chunk_index[t]], video frame t is control step t, and each policy_step maps to a contiguous block of control steps. Every episode passed these checks at collection time (n_control == 520 on failure = the step cap was hit).

Labels

Each episode carries a Gemini gemini-3.1-pro-preview description of its video and, for failures, a localized failure onset. Produced offline from rollout.mp4 + rollout.json only (no sim, no feature server); the exact prompt templates are in gemini_prompts.md / gemini_prompts.json and are also embedded verbatim in every labels.json under labeler.prompts.

Two passes on one shared video session:

1. Failure localizer (fork of Dan Lawson's liberox-evals) — a coarse pass on the full video produces the failure mode (8-way taxonomy), onset type (obvious_mistake / operator_intervention / timeout), onset time, and long + short reason / recovery / prevention strings; a second pass on a slowed ±window clip (1 clip-second = 1 rollout frame) refines the onset to a single frame. Only runs on episodes the simulator scored as failures; {} for successes.
2. 3-second keyword phrases — a second turn on the same session captions every consecutive 3 s window (2–6 word phrases, e.g. reaching for mug), given the failure summary as context but told not to copy it in.

labels.json per episode:

• rollout_id / scene_variant / task_id / task_file / instruction / success — enough to identify the episode standalone.
• labeler — {backend, model, refine, labeled_at, pipeline, prompts}.
• semantic_timeline — list of {segment_index, t_start_sec, t_end_sec, control_step_start/end, policy_step_start/end, phrase, description} (description is an alias of phrase).
• vlm_failure — the full localizer dict (mode, onset type/seconds/timestamp/ frame, coarse vs refined, confidence, justification, token usage, vlm_raw_response) plus vlm_failure_reason[_summary], vlm_recovery_action[_summary], vlm_prevention_action[_summary]; {} on success.
• failure_annotation — the VLM onset mapped onto the collection clocks: {failure_control_step, failure_sim_step, failure_policy_step, failure_chunk_index, first_post_failure_policy_step, failure_type, correction_action}.

labels.npz is the array-aligned copy for fast loading: sem_t_start, sem_t_end, sem_control_start, sem_control_end, sem_phrase, and fail_onset_frame, fail_onset_seconds, fail_mode, fail_reason, fail_reason_summary, fail_recovery, fail_recovery_summary, fail_prevention, fail_prevention_summary. example.md is a human-readable render of the same content.

Distribution. 500 labeled episodes, 3–9 timeline windows each. 110 have a localized failure: by mode — placement_or_insertion_failure 30, stuck_or_no_progress 27, grasp_failure 19, wrong_object_or_target 18, object_displacement 14, timeout_or_insufficient_progress 1, other 1; by onset type — obvious_mistake 79, operator_intervention 30, timeout 1.

Failure statistics

failure mode	n	% of failures	onset mean ± std (s)
placement / insertion	30	27.3%	14.3 ± 5.2
stuck / no progress	27	24.5%	12.4 ± 5.0
grasp failure	19	17.3%	12.6 ± 6.7
wrong object / target	18	16.4%	11.9 ± 3.0
object displacement	14	12.7%	13.3 ± 4.8
timeout / insuff. progress	1	0.9%	25.9 ± 0.0
other	1	0.9%	8.4 ± 0.0

Overall failure onset: 13.1 ± 5.2 s (median 12.5 s, range 1.2–25.9 s), out of 500 episodes total, 110 with a localized failure (22%).

The taxonomy defines 8 modes; unstable / dangerous behavior never occurs in this dataset (0 episodes) — listed for completeness, not omitted.

The keyword-phrase timeline and the free-text reason / recovery / prevention fields are open-ended language rather than categorical or numeric, so they aren't summarized as statistics here — see any episode's example.md for what they look like.

manifest.csv

One row per episode. Beyond the collection columns, the label pass adds labeled_captions, labeled_failure, labeler_model, and vlm_failure_onset_frame (blank for successes / unlabeled episodes).

Provenance

Collected with scripts/semantic_failure/ in the 12-Visual-Occlusion-Reasoning project against a local feature-serving pi05_libero server. replan_steps = 5, num_steps_wait = 10, seed 7, policy image size 224. Labels added afterward with the same scripts/semantic_failure/ pipeline and GEMINI_API_KEY only.