The robotwin_aloha_contact_recovery_dual_arm_v10 control profile separates
planning clearance from the physical execution boundary and adds a bounded
response to measured joint blocking. Its predecessor v9 assets remain frozen.
The v7 experiment/default bindings and live qualification are separate gates;
offline tests do not establish the physical effectiveness of retreat.
Clearance
The source-bound robot link spheres have radius 45 mm. A new planned path normally requires 95 mm between sphere centres; measured execution retains the 90 mm threshold. This is robot-versus-robot clearance, not an observed scene collision world.
When the action starts below the execution threshold, planning and local/global
execution use one immutable action-start measurement. Every stage must stay at
or above that distance minus 1 mm. The endpoint must increase the distance.
The reference is never reset at each waypoint or local iteration: many small
inward updates cannot accumulate a larger inward excursion. Receipt details
include escape_allowed, current_center_distance_m, the candidate distance
and both thresholds. An accepted escape can still be stopped by workspace,
joint, gross-motion or episode/terminal constraints.
The interval from 90 mm to 95 mm remains subject to normal planning clearance. The current implementation does not silently add another escape exception for that interval; its disposition remains a separate integration decision.
Blocking and feedback
On v10, the measured convergence loop counts actual physics samples for each
commanded arm against an unchanged joint target. Sixteen consecutive samples
with joint change below 0.001 rad and error above the selected convergence
tolerance stop the remaining path with not_reached / blocked. A changed
target or renewed joint movement resets the count. The 64-step waypoint limit
remains the fallback. Earlier profiles retain their previous wait behavior.
The receipt records blocked_at_stage, frozen_steps, progress and
stall_shape, plus the arms and measured joint error supporting the decision.
Progress is the zero-based blocked stage divided by the number of planned
stages. from_start means the first 10%; creeping requires three earlier
waypoints with waits above four steps; abrupt requires at least twelve
earlier waypoints whose first twelve waits are at most two steps. Other
prefixes retain stall_shape: null. These are measured execution descriptions;
they do not identify the contacting object or prove a contact force.
Recovery
If the preceding Move receipt was blocked, recover_motion first attempts a local
20 mm retreat opposite the recorded approach direction for affected selected
arms. Other selected arms hold their measured pose during that phase; gripper
directives remain unchanged. This is part of the same Supervisor-admitted
action and the same physical-step/terminal accounting.
For a global path, the direction comes from the last nonzero segment of the validated public-TCP path at or before the blocked stage. For a local action, it comes from the commanded local TCP displacement. The receipt preserves the source action and segment identity. This direction is not a collision normal and does not guarantee release.
Only a completed retreat permits global planning to the originally requested
previous/home goal. Both planners then use a fresh common measured snapshot.
The original action-start before state remains in the receipt. Both phases,
their control rows, their sample counts and their failures remain in canonical
evidence, including when the later global plan is rejected. A stopped or
terminal retreat does not proceed to global recovery.
A blocked stage zero or a pure rotation may supply no nonzero linear approach
segment. Missing or invalid direction evidence returns
blocked_recovery_direction_unavailable before any retreat write. A fresh
observation and an explicitly selected bounded local move remain available;
the Runtime does not invent a world axis or hidden contact information.
An automatic retreat outside the workspace similarly returns
blocked_recovery_retreat_outside_workspace without invoking global planning.
Ownership and verification
controllers/inter_arm_clearance.py owns the common immutable clearance rule;
controllers/motion_progress.py owns measured blocking and stall descriptions.
RoboTwin planning, global convergence and local Cartesian execution consume
these rules. v10
values live in its hash-bound control profile.
tests/ora/environment/backends/robotwin/test_robotwin_contact_recovery.py uses deterministic plants to verify
escape admission, cumulative inward rejection, execution-versus-planning
thresholds, exact 16-step blocking, terminal precedence, source-derived
retreat direction, fresh planning snapshots and preserved prefix evidence.
The pinned-source RoboTwin motion sweep and v7 behavioral triggers remain
required before production activation. No live motion was run to produce
these offline checks.
Public rejected-plan measurements
Recognized inter_arm_collision rejections retain their public explanation and
robot-only clearance measurements through the shared plan-feedback projection,
including single-plan failures and bounded candidate outcomes. stage_index
and link_pair locate the predicted rejection on the planned trajectory;
center distances and the current/required/planning/execution thresholds describe
the immutable admission decision. They are link-sphere center distances, not
an end-effector gap, measured collision or object geometry. The projection
neither changes the planning/execution margins nor permits an escape that the
controller rejected. Unknown or malformed numbers are omitted. The boundary
and Native OpenAI/Anthropic delivery regression live in
tests/ora/agent/session_core/test_plan_feedback.py.