evidence-driven-gripper — Claude Code skill: what it does, install
The evidence-driven-gripper Claude Code skill chooses contact actions from current observations and gripper capabilities. It checks object control before transport and revises the plan after a failed grasp.
MIT ★ 488 updated 2026-11-15
npx skills add Open-RUA/uni-robot-agent --skill evidence-driven-gripper -a claude-codeWhat evidence-driven-gripper does
The skill selects a contact operation from the instruction and current observations, such as pushing an object along its support or lifting it with a grasp. Before transport, it checks whether the hand controls the intended object. A successful tool call alone does not confirm motion or attachment.
It uses environment state, scene snapshots, and feedback from the hands. Snapshot coordinates remain useful while objects and supports stay put; after a slip, displacement, or uncertain release, affected features need fresh localization. The skill also assigns available hands roles such as carrying a load or adjusting its destination.
Its tool examples assume the uni-robot-agent interface, while actual capabilities depend on the robot environment and tools exposed in the session.
What it helps with
- choose between pressing, pushing, and grasping an object
- confirm a grasp before moving an object
- re-localize contact points after a slip or displacement
- assign separate roles to the available hands
Example prompts for the agent
- Check from the observations whether the hand holds the intended object before moving it.
- Push the object along its support instead of lifting it.
- After the grasp slips, re-localize the contact points and choose the next action.
How to install evidence-driven-gripper
1Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill evidence-driven-gripper -a claude-code2Manually: put the files in place
~/.claude/skills/general/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/general/
1Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill evidence-driven-gripper -a codex2Manually: put the files in place
.agents/skills/general/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill evidence-driven-gripper -a cursor2Manually: put the files in place
.cursor/skills/general/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Open-RUA/uni-robot-agent uni_robot_agent/resources/skills/general2Manually: put the files in place
.github/skills/general/SKILL.md1Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill evidence-driven-gripper -a gemini-cli2Platform CLI
gemini skills install https://github.com/Open-RUA/uni-robot-agent --path uni_robot_agent/resources/skills/generalManually: put the files in place
.devin/skills/general/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 evidence-driven-gripper
How do I install evidence-driven-gripper in Claude Code?
Run npx skills add Open-RUA/uni-robot-agent --skill evidence-driven-gripper -a claude-code, then use ~/.claude/skills/general/SKILL.md.
What does evidence-driven-gripper do?
It selects contact operations from instructions, observations, and available hand capabilities. It checks object control before transport and revisits the plan after a lost grasp.
What does evidence-driven-gripper need to work?
Its tool examples assume the uni-robot-agent interface. The available actions and feedback depend on the tools exposed in the current session.
Source text
SKILL.md · MIT
--- name: evidence-driven-gripper description: Choose and verify contact operations, placements and recoveries from current observations and the available gripper capabilities across tasks and environments. --- # Evidence-driven gripper Choose the operation from the current instruction and observations. Push or slide an object on its support when requested; a grasp-and-lift is not a substitute. Pressing needs visible contact and effect, not attachment. The acquisition and transport checks below apply when lifting is appropriate. ## Apply the current robot and runtime contract Use only the arms, tools and feedback capabilities exposed in this session. Two-hand guidance applies when two arms are available; single-arm tasks can use existing support or a permitted supported relay. Tool examples assume the stated uni-robot-agent interface, not universal robot geometry. Read the environment's robot manual and declared calibration when needed. Follow the current scheduling and physics-advancement contract before applying batching or waiting examples. ## Plan the available hands' roles before committing Before acquisition or a new contact operation, assess the available hands against the complete route. Assign the needed roles: carry, restrain, reposition the receiver, receive a transfer, or remain clear. Do not default to the first-used arm for every phase. Preserve any instructed hand assignment and existing support. When one hand holds the load and the receiver needs adjustment, assess whether the free hand can establish a usable destination without losing that hold. Read `general/gripper/supported-contact/SKILL.md` when the operation needs control of both objects or an assisting hand must adjust or restrain the receiver. Repeated single-arm route failures require reassessing hand assignments before another attempt; use `general/gripper/reachability-and-recovery/SKILL.md`. Coordination can be sequential. Keep a reliable single-arm route when existing support suffices; moving both arms adds no value without a useful second role. Coordinated Direct calls can move both hands together, but do not guarantee collision-free paths, synchronized arrival or a fixed relative pose. Read each arm's feedback and keep checkpoints where the next motion depends on new evidence. Use `general/gripper/pose-targeting/SKILL.md` for constructing pose targets. ## Keep a small evidence ledger For each relevant object, retain its identity, required relation, confirmation status and supporting observation. Track the operation actually needed; push, press and insert operations need not follow a pickup-to-placement sequence. Use `general/gripper/state-verification/SKILL.md` for unresolved states, evidence chains and completion checks. Use `general/gripper/contact-effect/SKILL.md` when contact effects need testing or correction. Read these when needed, not per call. Derive quantities, order and final relationships from the current instruction. ## Reuse coordinates only while their evidence remains valid Use existing evidence to choose the next action. Use view_env_state for missing visual evidence or hand state, and capture_scene when new coordinates are needed. Snapshots remain readable historical measurements after physical actions. Their coordinates remain useful while the objects, supports and relevant geometry remain unchanged; a readable snapshot need not describe the current scene. Re-localize affected features after slip, displacement or uncertain release; known lifts and clear waypoints can use measured hand state. A mask median is not automatically a grasp pose. Use the tool-selection skill when repeated queries fail to resolve the next action, rather than restarting scene mapping. ## Separate alignment, contact and attachment - Alignment means the hand reached the requested pose within its tolerance. Before closure, compare the actual contact-region pose with the grasp margin; pose_reached can still leave a thin object below the fingers. Use the arrival and slip cases in `general/gripper/grasp-geometry/SKILL.md` when evidence conflicts. - `call_ok` reports a valid call, not motion or object success. Read the completed action's stop reason, executed steps and measured state. Hand geometry and opening do not establish attachment; compare object motion and its support. ## Verify before committing to transport 1. Retain a pre-lift observation and the target's source location. Check visible material and clearance for both fingers before closing. A mask center may lie in a hole or between objects; it is not automatically a graspable point. 2. After closing, inspect execution feedback and the attached operating-side wrist image; use view_env_state if another view is needed. If visibly empty, reopen and realign. When lifting is intended and clearance permits, use a small verification lift. Compare source and hand views: did the intended object leave its support and follow the hand? A centered wrist image alone is insufficient. For constrained or fragile objects, choose a compatible probe. Dependent transport or insertion waits for confirmed control. 3. If attachment is confirmed, continue without extra numeric proof. If evidence conflicts, use a distinguishing view or compatible probe; missing segmentation alone does not prove loss. Use the state-verification skill for this decision. 4. Begin loaded carry with a short segment at the same orientation and confirm retention. Then follow the known clear route to the next meaningful checkpoint. Recheck changed load/support after substantial reorientation, obstacle contact or transfer, and check retention before release. On loss, reopen acquisition. For Direct verification and initial loaded transport, use short move(profile="precision") segments with the same orientation and a grip command compatible with the observed hold. A command does not guarantee retention. If the object shifts, revise the contact geometry and motion using `general/gripper/grasp-geometry/SKILL.md`. Do not enlarge motion increments just because the load is above the table; larger steps can cause slip. Use the exposed tool contract for limits and defaults. When actor_act is exposed, request one observable local phase and use its declared reviewer and chunk capabilities. Do not combine acquisition, transport and release or cross an unverified release boundary. The main agent inspects the result after the whole call; split phases into separate calls when a decision needs intervening evidence. Do not require a Direct setup sequence before every VLA action. ## Batch the known route, stop at unresolved decisions In parallel mode, submit already-determined independent queries or known route segments in one response, ending at the next unresolved decision. In particular, measure a manipuland and its receiver together when both queries are known. During a known clear motion, query an existing snapshot of an unaffected target before waiting. Read the batch examples in `general/gripper/tool-selection/SKILL.md` when choosing these boundaries. Preserve closure, attachment, support-transfer and release checks when later motion depends on their effects. Once confirmed, continue without another check for each small waypoint. Ordered execution does not certify route clearance. When a new locate query is already required after a planned physical action, submit that action followed by capture_scene in the same response. The shared physical FIFO captures after the action finishes. Receive its snapshot_id, then batch independent locate queries on that snapshot in the next response. If the need for localization depends on the action outcome, inspect the result first; for visual verification alone, reuse the action's returned images. Before a route crosses the shared workspace, inspect both wrists, forearms and held objects, even when only one arm will move. Identify the other arm's role: an empty idle arm can withdraw along a clear route; a hand supporting a container or holding an object cannot simply be parked away. Adjust the moving route or arrange another stable support without dropping that role. If motion stalls near the other arm, treat interference as a hypothesis, not a proven cause. Check actual pose and visible clearance. After changing the route or moving the obstructing arm, verify that change completed, then test the previously blocked segment before batching the rest. Merely announcing that the other arm is clear is insufficient. Shorter steps along the same blocked line can still stall and consume physics. Follow the runtime queue and barrier rules when resolving submitted work before recovery. A tail result does not prove an intermediate grasp worked. Once clearance and retention are established, resume the known route without making every free-space waypoint a separate decision. ## Preserve support and plan the whole route Use `general/gripper/reachability-and-recovery/SKILL.md` when arm choice or destination access is uncertain, `general/gripper/held-object-geometry/SKILL.md` when the load's actual extent determines clearance, and `general/gripper/moving-target-control/SKILL.md` when contact geometry moves. - Choose grasps compatible with the assigned roles and full route; plan any necessary relay before pickup. A failed pose does not prove all poses fail. - Reuse an already stable, reachable base instead of moving it for a uniform setup routine. Move it when the layout blocks later placements or withdrawal. Preserve useful racks and containers; follow any instructed hand-support role rather than assuming every container should be left on the table. - Plan clearance for the held object, both fingers and the forearms. A wrist or finger center above a rim does not prove the bottom of the load clears it. - For a hand-to-hand transfer, read general/gripper/handover/SKILL.md. Keep donor support until receiver closure, then verify independent retention after donor release and separation. Leave the receiving grasp and insertion features free. - Use a supported relay only when intermediate placement is allowed and direct transfer is infeasible. Check both arms' access before release; re-localize and verify the new grasp afterward. Laying a thin object flat can remove its useful grasp edge. A table relay does not fulfill an explicit hand-to-hand instruction. - Before placing, plan the empty-hand AND forearm withdrawal corridor. Release into stable support, withdraw clear and inspect the actual object and neighbors. If withdrawal disturbs the arrangement, reopen affected ledger entries. ## Match recovery to the observed failure | Feedback or observation | Next decision | |---|---| | max_steps with decreasing pose error | Continue toward the same absolute target if the route remains clear. | | IK failure, joint_jump_rejected or unsafe_motion | Read executed steps and current state; use `general/gripper/motion-recovery/SKILL.md` to choose a changed, checkable recovery. | | stalled or diverging | Check visible obstructions and actual tracking. Withdraw from contact before changing approach; do not merely raise the budget or push deeper. | | Hand closes but object stays at source | Check actual arrival error first. If alignment was sufficient, inspect material and two-finger clearance, then change grasp region or closing direction. | | Object lifts but slips in transport | Re-localize, then revise the grip/support relationship and route. Slower repetition of the same weak grasp is not necessarily an improvement. | | Object tips or neighbors move after placement | Reopen affected items and restore stable support before continuing nearby work. | Use `general/gripper/motion-recovery/SKILL.md` for partial execution, difficult reorientation and changed recovery routes. A previously reached pose does not certify the route back or clearance with a load. Separate substantial reorientation from carrying a fragile load, checking clearance and attachment. Tighter pose tracking does not prove object placement. ## Spend recovery effort on a changed hypothesis After unsuccessful attempts, distinguish the observed effect from a hypothesized cause. Change the relevant material region, closing axis, approach, support, route or arm based on that evidence. Avoid repeating an unchanged contact relationship without new evidence. If order is unconstrained, keep a difficult object pending and work on another reachable one. For a prescribed sequence, resolve the current item rather than skipping it. Use remaining_steps and costs observed in this episode to retain room for the remaining acquisition, transport, placement checks and safe withdrawal. Do not spend the remaining budget refining an already verified pickup. ## Inspect the physical result Check the instructed object relationships using current observations and still-valid process evidence. Occlusion alone does not invalidate a confirmed relation; an unverified relation remains unresolved. Recheck affected relations and their dependencies after a disturbance, using the state-verification skill. Use available action images or view_env_state for missing visual evidence, and calibrated geometry only when a correction needs new coordinates. Withdraw empty hands through clear space without disturbing the result. Complete the return required by the integration or task, using `general/gripper/return-to-start/SKILL.md`; do not add stricter pose requirements. For an instructed lift distance, compare the same object's position before and after pickup using visible geometry; hand displacement alone is insufficient, especially after slip or reorientation. Track the object rather than a rising corner caused only by rotation. Keep the measurement unresolved if occluded. Report visible progress and uncertainty faithfully during recovery; follow the system prompt's completion and stopping conditions. For robot frames, hand dimensions, jaw commands and camera calibration, use the current environment's published robot manual and capability description. Read available calibration for asset-specific numbers rather than guessing from a normalized drive value; metric openings require an exposed calibrated interface.
© Open-RUA · MIT · Open-RUA/uni-robot-agent