Robot disassembly learns from its own rollouts: PACER turns autonomous traces and human corrections into training weights, reaching 87/100 motherboard handoffs
Synopsis
In electronic-waste disassembly, demonstrations for vision-language-action (VLA) policies are scarce, and supervised fine-tuning alone struggles to cover the state shifts a deployed policy meets. PACER compiles autonomous rollouts, operator corrections and supplementary clean demonstrations into typed action-chunk rows, then weights each row by progress toward the declared target and by the trace's source and role, while keeping the native supervised action loss. On held-out hardware for UR5e motherboard disassembly across five components, the selected policy completed 87/100 approach-and-handoff trials, against 54/100 for uniformly weighted post-SFT and 79/100 for fixed-geometry reweighting.
Interpretation
Training data for a disassembly policy is no longer only clean demonstrations: the robot's own successful, near-miss and failed rollouts, plus operator corrections recorded from the failed state, enter the post-training set and supervise according to their process role. Prior post-SFT used supplementary clean demonstrations and corrections with uniform weights; the variant that also admits every eligible row uniformly reaches 64/100, while clean-only and correction-only variants fall below that baseline. On five components of UR5e motherboard disassembly, each method ran 20 trials per component on held-out hardware configurations, 100 trials in total; the selected PACER policy succeeded 87 times, fixed-geometry reweighting 79, uniform all-eligible 64, and uniformly weighted post-SFT 54.
Weights come from each action chunk's process evidence relative to the declared target rather than a binary success label: tool-center-point progress, final proximity, entry into the target region, displacement toward the target, and whether stop and handoff match the expected phase together set the row's share of the loss. Rows sharing a component, approach phase and process role are centered by median and scaled by a robust deviation, then mapped through a temperature, role multipliers and an upper clip into bounded weights, with protocol floors retaining clean and correction supervision. Eight predeclared weighting settings each fit a LoRA adapter on the training rows, and one is chosen by scores on configuration-disjoint validation rows plus safety audits; this run selected the setting that emphasizes terminal-region and stop-signal evidence.
Perspective
The findings rest on UR5e motherboard disassembly over five components: CPU, GPU, CPU fan, RAM and the connector block. The policy is evaluated as an action-chunk planner, scoring ends at the approach-and-handoff phase, and the downstream execution stack is fixed across methods. The held-out test uses two scene configurations per component with 10 trials each, 100 trials in total.
PACER relies on accurate trace metadata, including target geometry, process roles, operator labels, masks and provenance fields; robustness to broader metadata errors and corrupted training labels remains untested. Leading candidates are hard to separate under validation resampling, the eight candidates span less than a few thousandths of validation score, and calibration of that score against closed-loop hardware performance remains limited. The configuration comparisons and data-source ablations do not isolate the individual effects of evidence features, role multipliers, protocol floors or candidate selection.