Detect the bottleneck
Command and motion observations reveal that robot 2 no longer responds as predicted. The detector does not receive the hidden fault value.
← All demonstrations
Industrial robotics · Simulation
Keep the line in sync.
A slower robot becomes everyone's bottleneck. A08 models the two arms and conveyor together, then adjusts the line when one component changes.
The experiment
Five boxes. One unexpected slowdown.
The first arm marks each box. The second packs it. After box 2, the packing robot loses 30% of its action response. Without a new schedule, two boxes leave the conveyor unpacked.
3 / 5Original schedule
5 / 5A08 repair + replan
The robot is still slower.
The line works around it.
From component to system
A local repair. A coordinated response.
Repair one response
A08 estimates a local correction to robot 2’s action-response gain. An independent verifier checks prediction, scope and schedule feasibility before it is used.
Replan the line
The planner slows the conveyor to 82% of its original rate and retimes robot 1. Robot 2 stays degraded. The remaining boxes are packed at a sustainable cadence.
The result
Finish the batch, not just the motion.
Success means all five boxes are processed and packed. In this fixed scenario, waiting for the first missed box loses one item. Repairing the model and replanning the line completes the batch.
| Route | Packed | Mission |
|---|---|---|
| Unchanged model and schedule | 3 / 5 | Failed |
| Slow down after the first miss | 4 / 5 | Failed |
| A08 repair and line replan | 5 / 5 | Completed |
How the demonstration was built
The action-conditioned line model includes a nominal response backbone fitted on 108 cycles. The candidate changes only robot 2's response gain, estimated from command and progress observations. The exact fault value is reserved for the reference route.
The unchanged nominal scene and the exact-reference route each pack 5 of 5 boxes. These are feasibility checks, not additional A08 recovery trials.
ManiSkill and SAPIEN provide the articulated robot scene. Arm trajectories are authored and box transport is kinematic. This demonstrates model repair and schedule coordination, not learned grasping or contact robustness.
Simulation only
A bounded engineering demonstration, not production, safety or physical-hardware validation.
The fault remains
A08 changes the model and operating schedule. It does not restore the robot's original motor response.
Independent work
ManiSkill, SAPIEN and xArm6 identify the simulation tools and assets. No manufacturer partnership is implied.
Design partnership
What happens when one machine slows your line?
Bring a production cell, its telemetry and one repeatable failure mode. Let's define a measurable pilot.