corvinus-labs/regen_s2_200_seed29700

regens2200_seed29700 — S2 (appearance) staged-regen batch

200 successful tube-rack insertion episodes from Isaac Sim with a myCobot 320 M5, generated as the S2 config of the four-way staged regen. S2 is S1b plus full appearance randomization and nothing else, so S1b vs S2 at equal N measures what appearance DR costs and buys.

Companions: S1 (rotation only), S1b (+ recovery).

Contents

• —200 episodes, 200/200 success, 17,939 frames, 4.8 GB
• —63–159 steps per episode (mean 89.7), 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 S2 variable: full appearance DR

appearance_mode: full in all 200 episodes — table material/texture/colour, rack colour/metallic/roughness, tube colour/opacity/specular, liquid (present in 151/200), and light count (1–3 keys plus dome). Everything else is held at S1b's settings.

Still fixed, and this is what separates S2 from S3: one rack (rack2), one tube model (50ml_tube), no clutter, no bore tubes. So an S2 result is attributable to appearance alone; S3 changes geometry as well.

Recovery data (inherited from S1b)

Mid-episode shoves, perturb_count [1,2] — 1 shove in 102 episodes, 2 in 98. 3,086 of 17,939 chunks (17.2%) cross a shove and are clamped. Every record carries valid_horizon and crosses_perturbation; consumers must honour them or a clamped tail trains as a genuine label.

Generation

• —Task variant tube_rack_online_insert_v0, embodiment mycobot_320_m5
• —Seeds 29700–29916, 200 kept of 217 attempts — disjoint from S1 (28200–28417), S1b (28700–28925), S3 (30700–30936), v3 (1000–1499), franka (1000–1049), pilot (8000–8050)
• —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

Provenance caveat

batch_manifest.json records commit bec5b87, which is 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). It is not resolvable from the remote. The generator code it carries is that branch's sim/ tree; the equivalent becomes resolvable once the generator branch merges to main. git_dirty: true refers only to infra/sim/isaac_rack.py and infra/sim/rack_logic.py — eval-harness files that take no part in generation.

Known caveats

28.4% of frames are the post-success hold (hold_steps_after_success 25, 5,094 frames near zero motion). Relevant if you score translation against a zero-motion baseline; metadata.json success_step marks the boundary.

Task naming. The staged-regen note calls for tube_rack_align_v0 with expert approach_gain 0.25; this batch is stamped tube_rack_online_insert_v0, which is what --regen-config produces. approach_gain does not exist on the online-insertion path. Same path as S1/S1b, so the cross-config comparisons are unaffected.

Action chunks

vla_chunk_anchored_v3 — H=16, action_dim 9, camera frame: a_i = R_cam(t)ᵀ (p_cam(t+i) − p_cam(t)), 6D continuous rotation, camera axes +x right / +y up / −z forward (USD optical), padding repeat_final_pose. 17,939 records, 1:1 with frames.