Interpret RoboDojo motion and recovery
Use general/gripper/pose-targeting/SKILL.md to construct pose targets and
general/gripper/motion-recovery/SKILL.md to choose a changed recovery from
observed evidence. The details below describe this backend, not other robots.
Waypoint planning and temporary goals
Profiles set waypoint limits and final arrival tolerances; they add no approach
height. RoboDojo keeps strict native waypoint IK (1 mm / 0.02 rad). After a
planning rejection, one recovery search per move may choose a temporary execution
goal within 80% of the original final position and orientation tolerances.
The requested target remains unchanged for measured arrival checks. Read
goal_recovery and per-arm execution_target_position/execution_target_quaternion
to distinguish the route's temporary endpoint from the original target_pose.
There is no automatic return to the exact goal center or coarse-to-fine phase.
Choose tolerances that the intended operation permits before submitting move.
Recovery diagnostics
Per-arm recovery.rejected records stages for finally rejected candidates.
native_not_converged and native_constraint_rejected are distinct; a candidate
can have both. collision_or_native_constraint means a sampled joint connection
failed a native constraint check, not necessarily that endpoint IK failed.
goal_recovery records the optional once-per-move final-region search. A found
candidate is only a temporary execution goal: subsequent waypoints still need
IK and measured arrival must satisfy the original goal's tolerances. There is
no automatic tightening phase. Inspect actual feedback before dependent actions.
Execution feedback and limits
A successful tool call does not prove physical arrival or task success. Inspect
feedback.pose_reached, actual position/orientation, executed steps and stop
reason. grasp or release does not prove object attachment or detachment.
Use measured state and images to verify the intended effect.
The current native IK includes modelled self/table constraints. Other action routes do not necessarily have identical checks. Endpoint checks and joint jump guards do not establish collision-free swept motion; task objects, the other arm and carried objects are not automatically represented by the public robot geometry. Plan short observable approaches and check clearances.
This interface does not add a timed-trajectory guarantee. Control steps consume native episode budget; model thinking and network delay are separate wall time. On a partial or uncertain action, inspect returned state/status instead of replaying the same physical request.
IK recovery goals and counts
RoboDojo keeps each native waypoint solve at 1 mm / 0.02 rad. After a rejected
waypoint, move may search once for a temporary endpoint inside 80% of the original
final position/orientation tolerances. target_pose remains the requested goal;
execution_target_position and execution_target_quaternion identify the
possibly adjusted endpoint. goal_recovery reports that search, not physical
arrival. Actual pose must still satisfy the original tolerances; there is no
automatic return to the center or coarse-to-fine phase.
Primitive per-arm recovery.stats and move's aggregate ik_statistics count:
ik_requests: one arm's strict waypoint planning request, including held arms when actually solved; arms skipped after an earlier rejection are not counted.ik_request_failures: requests still rejected after candidate recovery.solver_calls/seeds_attempted: 32 then 128 means two calls and 160 seeds.candidates_returned: distinct seed results, with the true top1/topN alias removed.candidates_checked/candidates_rejected: candidates examined / finally rejected. Raw and legal whole-turn representations are branches of one candidate.
seed_budget retains its existing meaning: the largest single solve's seed
budget, not the sum. Different seeds with equal joint vectors remain distinct.
rejected counts each finally rejected candidate at most once per stage; stages
can overlap, so their sum is not a count of failures. Raw top1 convergence, native
feasibility and pose errors are reported separately from its combined success.
ik_candidate_rejected covers validation/connection failures and mixed stages;
none of these finite searches proves the requested pose mathematically impossible.
The waypoint IK rejection rate is ik_request_failures / ik_requests when the
denominator is nonzero. Region-search calls are separately recorded under
goal_recovery.diagnostics and are excluded from that fraction. Do not sum move
aggregates and their primitive receipts together. Across logs, identify each
primitive request and arm once. ik_statistics=null means this backend provided
no such counters, not that zero IK requests failed.
Recovery and episode stopping
IK failure and joint-jump rejection are recoverable planning feedback, not task failure or an environment failure preventing interaction. While the task remains incomplete, interaction is available and budget remains, continue recovery. Neither one rejection nor repeated rejections justify finish, ending the turn, or declaring the task impossible. Abandon an ineffective route, not the goal.