return-to-start — Claude Code skill: what it does, how to install
return-to-start — a skill for Claude Code and Codex guides robotic arms back to their measured starting poses. Verifies position and orientation tolerances before ending the movement.
MIT ★ 488 updated 2026-11-15
npx skills add Open-RUA/uni-robot-agent --skill return-to-start -a claude-codeWhat return-to-start does
The skill handles returning empty robotic arms to their initial measured positions and orientations. It retrieves the reference pose via view_env_state at step=0 or from the initial environment state. Motion executes through move using [x,y,z,roll,pitch,yaw] coordinates derived from the original position and quaternion at the registered grasp site.
Returning is treated as a physical motion within the current episode, prohibiting environment resets, zero joint angle assumptions, or recalibrations. The arm must clear manipulated objects after confirming release. If paths conflict, arms return sequentially using intermediate waypoints rather than simultaneous trajectories.
The pose_reached flag only signals completion of the command, not alignment with the true starting posture. The procedure requires verifying both position and orientation against control_pose under identical coordinate frames, consulting general/gripper/pose-targeting/SKILL.md if grasp site semantics require resolution.
What it helps with
- return an empty robotic arm to its measured starting coordinates and quaternion
- verify stable object release prior to executing arm withdrawal
- prevent arm collisions by sequencing trajectories or inserting intermediate poses
- validate final grasp site pose against initial control_pose tolerances
Example prompts for the agent
- Return the idle arm to its starting pose using view_env_state step=0.
- Verify the gripper reached the measured initial orientation before closing the task.
- Plan a collision-free return path for both arms and check pose tolerances.
How to install return-to-start
1Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill return-to-start -a claude-code2Manually: put the files in place
~/.claude/skills/return-to-start/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/return-to-start/
1Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill return-to-start -a codex2Manually: put the files in place
.agents/skills/return-to-start/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill return-to-start -a cursor2Manually: put the files in place
.cursor/skills/return-to-start/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Open-RUA/uni-robot-agent uni_robot_agent/resources/skills/general/gripper/return-to-start2Manually: put the files in place
.github/skills/return-to-start/SKILL.md1Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill return-to-start -a gemini-cli2Platform CLI
gemini skills install https://github.com/Open-RUA/uni-robot-agent --path uni_robot_agent/resources/skills/general/gripper/return-to-startManually: put the files in place
.devin/skills/return-to-start/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about return-to-start
How to install return-to-start in Claude Code?
Run npx skills add Open-RUA/uni-robot-agent --skill return-to-start -a claude-code. The skill configuration saves to ~/.claude/skills/return-to-start/SKILL.md.
What does return-to-start do?
It directs the agent to safely return robotic arms to initial measured coordinates, avoiding collisions and confirming tolerances.
What is required for return-to-start to work?
The robotic environment must provide view_env_state and move tool interfaces supporting pose targeting and quaternion conversions.
Source text
SKILL.md · MIT
--- name: return-to-start description: Return empty arms to their measured initial poses and verify the finishing posture. Use before returning an arm or checking whether it has returned. --- # Return to the measured initial pose The integration and task instructions determine when return is required and its acceptance tolerances. Restore both position and orientation within those tolerances. Moving clear of the objects, reaching a familiar XYZ, or receiving pose_reached does not establish that the arm has returned. ## Recover the correct target When recorded initial observations are exposed, use view_env_state with step=0 to recover each arm's initial measured pose; otherwise retain the initial measured state supplied by the environment. Keep each arm's own position and quaternion together. Use move with that arm's initial position and its quaternion converted to fixed-axis XYZ Euler radians in pose=[x,y,z,roll,pitch,yaw], both measured at its registered grasp site. Choose the motion profile and tolerances from the tool contract. Do not retain a top-down grasp orientation when the initial orientation differs. Use the environment's declared return reference and acceptance criteria. If its scoring reference differs from the control point, use the published relationship rather than transferring a position tolerance between frames. Use `general/gripper/pose-targeting/SKILL.md` when the grasp site or rotation semantics are unclear. Returning is physical motion within the current episode. Do not reset the environment, recalibrate the robot, or substitute assumed zero joint angles. ## Plan a clear return First establish stable support for any held object and verify release. Preserve contact when the task still requires the arm to support or hold something; do not return that arm prematurely. Withdraw clear of the object before a large translation or reorientation. Check the swept path of the fingers and both arms, not just the endpoints. Use intermediate poses where needed. Coordinate both arms only when their routes do not interfere; otherwise return them in a suitable order. ## Verify the complete pose Resolve submitted motion before inspecting its final state. Compare the measured grasp-site position AND orientation with the initial control_pose, using the same reference frame. Check all available arms, including any that may have moved unintentionally. pose_reached refers to the target supplied to that call. Verify that the supplied target was the intended initial pose before relying on that flag. If either position or orientation is outside the required tolerance, correct the observed mismatch along a clear path and inspect the result. Check that withdrawal left the objects supported and undisturbed. A successful return establishes the finishing posture, not task completion.
© Open-RUA · MIT · Open-RUA/uni-robot-agent