datagen-stack-safetylabeled-raw

Raw stack_retrieve trajectories (unstack three objects from a target object onto a pile beside it, then retrieve the target into a goal region) collected with the LTL safety filter off, every trajectory labeled safe or unsafe.

Purpose: built for a standalone, temporary experiment; not part of the main data release. For general SFT use datagen-stack-v1-joint-5cam.

Contents

Natural collection: 2,148 attempts → 1,120 successes, of which 273 unsafe (24.4% of successes). Violated clauses: target_upright 242, stack_upright 222, no_stack_dropped 4 (an episode can violate several). The pool holds 1,133 trajectories: 1,120 from the natural collection (40 successes per task) plus 13 from 27 later attempts on tasks 0004, 0006, 0012, 0015, 0017, 0022 and 0023 that collected one more unsafe success each (7 unsafe, 6 safe); those 27 attempts are the last rows of the tasks' attempts.jsonl (2,175 rows in total).

Sibling repos built from the same pool: datagen-stack-allsafe40-joint-5cam · datagen-stack-unsafe25-joint-5cam · datagen-stack-unsafe25-safetytag-joint-5cam · raw pool datagen-stack-safetylabeled-raw.

Collection

Demonstrations come from the same cuRobo planner pipeline, bench tasks and cameras as datagen-stack-v1-joint-5cam, with one difference: the LTL safety monitor ran on every step but did not reject unsafe attempts. Every successful attempt was kept and labeled safe or unsafe by the monitor (fixed draw stream starting at draw index 1000). An episode is unsafe if it violates at least one of the family's safety clauses:

Clause	LTLf	Meaning
no_stack_dropped	G (!any_stack_dropped)	no stacked object falls to the floor
stack_upright	G (all_stack_upright)	all stacked objects stay upright
target_not_dropped	G (!target_dropped)	the target never falls to the floor
target_upright	G (target_upright)	the target stays upright

Layout

task_XXXX/
  attempts.jsonl      every attempt, failures included (fail_stage, seed, draw_index, safety label)
  labels.jsonl        one row per kept trajectory: safe, violated_constraints, violation_frame
  _summary.json       per-task counts (n_attempts, n_safe, n_unsafe)
  traj_YYY/
    meta.json         prompt, seed, draw_index, grasp, safety label
    traj.hdf5         state (T,8), actions (T,8), actions_commanded (T,8), states (T,121) sim state, datagen_info
    safety_trace.jsonl  per-step atomic propositions and monitor state
    image_opposite.mp4 image_left.mp4 image_right.mp4 image_left_shoulder.mp4 wrist_image.mp4  (30 fps, 256×256)

safety_trace.jsonl: row 0 holds the initial propositions; rows 1..T correspond to frames 0..T−1; the last 45 rows are a rest settle after the episode that is not recorded in the videos.

License

Released under CC BY-NC 4.0 (see LICENSE). The generated artifacts — joint state/action trajectories, rendered demonstration videos, and language prompts — are © 2026 IDEAS Lab, Northwestern University.

The underlying scene and object assets shown in the rendered videos come from BEHAVIOR-1K and remain subject to its license; this dataset does not redistribute them as assets.