mars-black/mars_black_grasp_yellow_cube_v1

Mars_black — grasp_yellow_cube_v1

Real-robot demo dataset from mars-the-40th (Innate Maurice, Jetson Orin Nano Super). Brain + Cerebellum + Evaluation agent stack (docs.innate.bot). Logged via Claude Code agents.

Task

grasp_yellow_cube_v1 — given a workspace of foam cubes in front of the arm, perceive a yellow cube via RGB segmentation, plan a top-down grasp via KDL IK + arm-camera fine-tune, close the gripper, lift, then re-pick the next-color cube. Two rounds executed on hardware.

Result overview

Phase	Color	Outcome	Bag size	Notes
A (warmup)	yellow	partial — not archived as bag	—	torque/home calibration, used to seed phase_b
B	yellow	grasped + lifted	~31 MB	clean run, 15 audit steps
C	purple (re-pick)	grasped + lifted	~706 MB	post-release, re-perceive, retry policy, 66 audit steps
Finale	release purple + home	EE 4.2 cm from home target (within 5 cm tol), head=0	—	see finale/

Each phase has end-effector + arm-camera keyframes at every audit point. The bag is recorded across the entire grasp window and uploaded here.

Videos

Compressed-image topics from each phase's rosbag2 have been re-encoded to mp4 (H.264-ish, fourcc mp4v) with a top-left watermark (phase | camera | t=bag_relative_seconds).

Phase	Camera	File	Frames	FPS (median)	Duration	Resolution	Direct link
B	arm_camera (link5)	phase_b/videos/phase_b_arm.mp4	442	5.09	87.58 s	640 x 480	download
C	arm_camera (link5)	phase_c/videos/phase_c_arm.mp4	672	5.08	133.31 s	640 x 480	download
C	main_camera/left (head)	phase_c/videos/phase_c_main.mp4	954	7.14	133.27 s	640 x 480	download

Phase B bag did not include the main-camera topic, so only the arm-camera mp4 exists for phase B. Each video directory also contains _thumb_first.jpg / _thumb_mid.jpg / _thumb_last.jpg plus a manifest.json with the exact bag timestamps and topic name. Source bags under phase_b/bag/ and phase_c/bag/ are unmodified.

Directory layout

phase_b/
  bag/phase_b_bag/         rosbag2 (db3 + metadata.yaml)
  checkpoints/             keyframe JPGs at audit points (CP0_..CP6_)
  videos/                  phase_b_arm.mp4 + thumbs + manifest.json
  logs/                    text logs
  steps_log.jsonl          one JSON record per audited step (input/output/ok/duration/ts)
  run_phase_b.py           executor with audit-keyframe insertion

phase_c/
  bag/phase_c_bag/         rosbag2 (db3 + metadata.yaml, larger window)
  checkpoints/             P/Q/R-section keyframes
  videos/                  phase_c_arm.mp4 + phase_c_main.mp4 + thumbs + manifest.json
  logs/                    text logs
  perception/              cube-detection overlays per attempt
  steps_log.jsonl          one JSON record per audited step
  run_phase_c.py           executor (release + re-perceive + re-grasp w/ retry)

finale/
  K1_pre_release.jpg       arm camera before opening jaws
  K2_post_release.jpg      arm camera after dropping cube
  arm_after_home.jpg       arm camera after returning to home pose
  steps_log.jsonl
  summary.json
  run_finale.py

Field & unit notes

• position_m — base_link frame, meters.
• orientation_quat — wxyz internal but reported xyzw in JSON.
• pitch — radians; default pre-grasp pitch ≈ 1.4451 rad (~82.8°).
• joint angles — radians.
• effort_pct is misnamed. Real values range from −41 to +1103 during gripper close (see phase_c/steps_log.jsonl). The underlying value is the raw load reported by the servo, not a 0-100% normalized value. Treat as a raw integer; the threshold used to detect a successful grasp in this dataset is abs(effort) > ~700. To be fixed in the next iteration of perception-joint-states.
• Audit keyframes (K* / CP*) are always paired with EE perception logging within the same second.

Hardware

• Robot: Innate Maurice (mars-the-40th), Jetson Orin Nano Super.
• Arm: 6-DoF, KDL IK (imprecise per company), joint6 = gripper, reach ~0.27 m.
• Cameras: main_camera (head), arm_camera (link5, 25° downward URDF pitch, ~60° world pitch in pre-grasp; NO camera_info published — naive pixel-to-metres scaling fails).
• Surface: soft foam cubes on a soft tabletop, safe for a ~10 cm release drop.

Reproducing

run_phase_*.py are the exact executors. They depend on the Mars_black skill set at /home/jetson1/.claude/skills/ (not uploaded — controller / perception code is robot-specific). They are included as documentation of the audit-keyframe insertion pattern used by the evaluation-agent.

Citation

Internal demo run. If you use this data please cite:

Mars_black evaluation-agent, grasp_yellow_cube_v1, Innate Maurice (mars-the-40th), 2026-05-21.