Gallery · RoboTwin

stack blocks three, seed 100003

Task: “Bring red block, green block, blue block to the table's center and stack them: blue block on green block, green block on red block.” DeepSeek-V4.1-Flash, code mode: 7 decisions, 39 motions, solved.

2× · 60 s of motion
  1. 1 of 7 · step_red1.py
  2. 2 of 7 · step_red2.py
  3. 3 of 7 · step_green1.py
  4. 4 of 7 · step_red3.py
  5. 5 of 7 · step_red4.py
  6. 6 of 7 · step_green2.py
  7. 7 of 7 · step_blue1.py
  1. grasp_execute
  2. grasp_execute
  3. grasp_execute
  4. grasp_execute
  5. grasp_execute
  6. move_to
  7. move_to
  8. set_gripper
  9. move_to
  10. recover_pose
  11. grasp_execute
  12. grasp_execute
  13. grasp_execute
  14. grasp_execute
  15. grasp_execute
  16. move_to
  17. grasp_execute
  18. grasp_execute
  19. grasp_execute
  20. grasp_execute
  21. grasp_execute
  22. move_to
  23. move_pose
  24. set_gripper
  25. move_to
  26. move_pose
  27. move_pose
  28. set_gripper
  29. move_pose
  30. move_to
  31. recover_pose
  32. grasp_execute
  33. grasp_execute
  34. grasp_execute
  35. grasp_execute
  36. grasp_execute
  37. move_pose
  38. move_pose
  39. set_gripper
Decisions

The recording is the robot's own cameras at 2×; the time between motions, while the model thinks, is not filmed. The 3D replay is the camera's picture in depth after each decision, with the paths it planned, to turn and step through.