corvinus-labs/regen_s3_200_seed30700

regens3200_seed30700 — S3 (variety) staged-regen batch

200 successful tube-rack insertion episodes from Isaac Sim with a myCobot 320 M5, generated as the S3 config of the four-way staged regen. S3 is S2 plus scene variety: five rack configurations, two tube models, clutter and bore tubes.

Companions: S1, S1b, S2.

Contents

• —200 episodes, 200/200 success, 17,793 frames, 4.4 GB
• —69–138 steps per episode (mean 89.0), one 640×480 RGB PNG per step
• —Per episode: step_NNNN.png, steps.jsonl, steps_chunked.jsonl, chunk_meta.json, trial.json, metadata.json, deltas.npy
• —batch_manifest.json at the root

The S3 variable: geometry and clutter

Attribution warning, stated in the staged-regen note itself: rack variety changes appearance and geometry at once — bore layout and rim height differ between rack configs — so an S3 failure is not attributable between the two. The note accepts this and calls S3 the least critical of the four configs. Do not read an S3 result as an appearance finding; that is what S2 is for.

Recovery data (inherited from S1b)

perturb_count [1,2] — 1 shove in 95 episodes, 2 in 105. 3,102 of 17,793 chunks (17.4%) cross a shove and are clamped. Every record carries valid_horizon and crosses_perturbation; consumers must honour them.

Generation

• —Task variant tube_rack_online_insert_v0, embodiment mycobot_320_m5
• —Seeds 30700–30936, 200 kept of 237 attempts — disjoint from S1 (28200–28417), S1b (28700–28925), S2 (29700–29916), v3 (1000–1499), franka (1000–1049), pilot (8000–8050). The higher attempt count (237 vs S2's 217) is the cost of the extra geometry: more draws are rejected when bore layout and rim height vary.
• —Expert: online branch-continuous Cartesian IK, step_size_m 0.0055, orientation_step_rad 0.0873; executor curobo-pd-v0, cuRobo 0.8.0.post1.dev36
• —Camera: tool_camera, 640×480, fx 480.599 fy 484.479 cx 316.133 cy 228.148
• —Liquid present in 154/200 episodes

Provenance caveat

batch_manifest.json records commit b463fa1, on a local integration branch that was never pushed (main + the generator branch sacchin/mycobot-dr-bench-and-random-start + in-progress eval-scene work), so it is not resolvable from the remote. git_dirty: true refers only to infra/sim/isaac_run.sh, an eval-harness launcher that takes no part in generation.

Known caveats

28.7% of frames are the post-success hold (5,104 frames near zero motion). success_step in metadata.json marks the boundary.

Task naming. Stamped tube_rack_online_insert_v0 rather than the note's tube_rack_align_v0; same path as S1/S1b/S2, so cross-config comparisons are unaffected.

Action chunks

vla_chunk_anchored_v3 — H=16, action_dim 9, camera frame, 6D continuous rotation, padding repeat_final_pose. 17,793 records, 1:1 with frames.