PACER: Process-Aware VLA Policy Learning from Human Corrections for Robotic Disassembly

• Robotic disassembly remains largely manual because workers adapt to product variation, whereas many robotic systems are designed for known products and fixed procedures.Manual disassembly is labor-intensive and hard to scale.Source: §1 · p. 1
• SFT data are scarce and narrowly covered, while rollouts and corrections capture unfamiliar states and recovery actions that clean trajectories may not cover.The paper investigates learning from these experiences through SFT rather than requiring a unified RL approach.Source: §1 · p. 2
• 87/100 held-out approach-and-handoff trials succeed with PACER, compared with 79/100 for fixed-geometry reweighting and 54/100 for vanilla post-SFT.The evaluation covers real UR5e motherboard disassembly across five component classes.Source: §1 · p. 3
• PACER weights valid action chunks using progress toward the target and each trace’s source and outcome, while retaining the native supervised action loss.It learns from autonomous rollouts, human corrections, and supplementary clean demonstrations.Source: §1 · pp. 2–3

• Most VLA models are trained for daily-life tabletop tasks, unlike industrial disassembly with product variation, limited demonstrations, and contact-rich execution.Related resources include robot datasets such as DROID [24] and the lifelong-learning benchmark LIBERO [25].Source: §2 · p. 3
• Advantage-weighted regression and AWAC [16] [28] weight behavior-cloning losses by estimated advantages, while L2RW and Meta-Weight-Net [29] [30] use validation feedback to learn weights.Source: §2 · p. 3
• Online VLA post-training uses policy gradients with sparse success rewards [32] [33] [34], and RECAP and SOP [36] fuse teleoperated interventions with autonomous rollouts.Source: §2 · p. 3
• Correction-learning methods prioritize interventions or turn them into preferences, while reward-based and offline approaches rely on reward supervision or value estimates [37] [38] [39] [40] [41] [14] [43] [44].Flow-matching VLA heads complicate likelihood-based policy-gradient updates [12].Source: §2 · pp. 3–4

Policy selection

• A fixed analytic validator scores predicted chunks on logged validation observations, applies safety and validity audits, and ranks candidates by a component-balanced validation score.If no candidate passes the audits, PACER selects the supervised reference policy; otherwise, it evaluates the selected policy on held-out hardware tasks.Source: §3 · pp. 4, 6, 7

Data collection

• PACER collects autonomous rollouts, corrections from reached states after eligible unsuccessful attempts, and supplementary clean demonstrations under a fixed coverage protocol.Rows from the same case group are kept together in the fixed training-validation split, and hardware test scenes are reserved for final evaluation.Source: §3 · p. 4

Row weighting

• An eligibility gate excludes invalid or unsafe imitation targets, while process evidence describes progress, proximity, terminal-region entry, direction, stop/handoff agreement, operator outcome, and provenance.Trace roles are assigned from provenance and outcome labels before candidate weights are computed.Source: §3 · p. 5
• PACER centers process scores within component-phase-role groups, combines them with role multipliers and protocol floors, and clips final row weights to [0, wmax].Temperature controls emphasis on above-median evidence; floors retain clean and correction supervision.Source: §3 · p. 6

Weighted training

• Each candidate uses offline fixed row weights and valid-horizon masks to scale the native action loss, leaving action targets unchanged.Fresh LoRA adapters are fitted from the shared supervised reference checkpoint under the same post-training recipe.Source: §3 · p. 6

Setup

• The UR5e benchmark covers CPU, GPU, CPU fan, RAM, and connector-block disassembly, using a reference policy fine-tuned on 710 teleoperated trajectories.RAM and connector-block demonstrations have lower per-configuration coverage than the other three components.Source: §4 · p. 7
• Each method is tested on 100 held-out trials: 10 per configuration across two configurations for each of five components.The configurations include one close to the training distribution and one out of distribution.Source: §4 · p. 7, Table 1 · p. 8

Methods

• PACER trains eight predeclared weighting settings and selects one using validation, while each baseline trains one policy under a fixed setting.Vanilla post-SFT uses supplementary clean demonstrations and corrections with uniform weights; Uniform all-eligible also includes eligible autonomous traces.Source: §4 · p. 7

Controlled OOD comparison

• Table 2 compares policies on two distinct OOD configurations, with mild and strong shifts and 50 trials per policy per configuration.Its paired bootstrap intervals report selected PACER minus each comparison row.Source: Table 2 · p. 8

Results

• 87/100 trials succeed with PACER, versus 79/100 for Fixed geometry, 64/100 for Uniform all-eligible, and 54/100 for vanilla post-SFT.The selected setting emphasizes terminal-region and stop-signal evidence.Source: §4 · p. 8, Table 1 · p. 8

Ablation design

• The ablations compare three source removals and fixed, preset, uniform, and role-only weighting alternatives.No clean removes supplementary clean demonstrations; No correction removes human corrections; No outcome removes autonomous rollout data.Source: §5 · p. 9

Data-source ablations

• Removing any post-training data source reduces success, with the largest decline when corrections are removed.The correction-only policy records no successful approaches from nominal starts; removing autonomous outcomes also lowers success despite retaining both forms of human supervision.Source: §5 · p. 9

Weighting comparisons

• Among policies using all three data sources, selected PACER outperforms the fixed weighting alternatives, and its paired-bootstrap interval versus Fixed geometry lies entirely above zero.The comparison uses policies trained from the same reference policy with the same nominal training budget.Source: §5 · p. 9

Data-source ablations

• Corrections show recovery and refinement from policy-induced states, while supplementary clean demonstrations show normal progress from nominal starts.The No outcome policy performs worse than selected PACER, indicating added value from autonomous execution experience beyond clean and corrective demonstrations.Source: §5 · p. 9

Metadata

• PACER relies on accurate target geometry, process roles, operator labels, masks, and provenance metadata.Robustness to broader metadata errors and corrupted training labels remains untested.Source: §6 · p. 9

Evaluation scope

• Evaluation is limited to motherboard disassembly on one robot platform and one VLA model, with no evaluation yet on other domains or standard manipulation benchmarks.Hardware results have not yet guided iterative refinement of PACER.Source: §6 · p. 9

Validation and attribution

• Validation resampling leaves leading candidates difficult to distinguish, and calibration of Jval against closed-loop hardware performance remains limited.The experiments also do not isolate individual evidence features, role multipliers, protocol floors, or candidate selection.Source: §6 · p. 9

7 Conclusion

• PACER improves approach-and-handoff success over vanilla SFT post-training and fixed-geometry reweighting on the evaluated motherboard tasks.The paper concludes that process-aware use of heterogeneous experience can make SFT more effective when disassembly demonstrations are limited.Source: §7 · p. 9

A.1 Trial accounting and configuration terminology

• The held-out benchmark uses two scene configurations per component, one near and one outside the training distribution, with 10 trials per configuration.Test configurations are fixed before candidate selection and excluded from the demonstration and post-training data.Source: §A.1 · p. 13

A.2 Hardware platform and success criterion

• The UR5e platform evaluates five motherboard components, and scored trials stop before downstream skill execution.The operator judges whether the fixed downstream skill can execute reliably from the final pose and visible state, using the same handoff rubric for every method.Source: §A.2 · p. 13

A.3 Component tasks, targets, and thresholds

• CPU targets use a 0.005 m position tolerance, other components use 0.010 m, and all share a 0.35 rad orientation tolerance.The proximity normalizer is twice the component’s position tolerance; canonical prompts route components to yellow, blue, or green zones.Source: §A.3 · p. 13

A.4 Closed-vocabulary outcome labels and labeling protocol

• Trials receive one of five outcome labels, and strict success requires the success label without timeout, unsafe abort, or intervention before valid handoff.One operator labels each recorded final state using the shared rubric; labels are fixed before aggregate method performance is computed.Source: §A.4 · pp. 13–14

B.3 Rollouts and Corrections

• The reference policy is fine-tuned on 710 teleoperated trajectories, then collects 70 autonomous rollouts per component across seven configurations.Eligible near misses and failures are immediately followed by teleoperated corrections from the reached state, stored separately but associated with the attempt.Source: §B.2 · p. 15, §B.3 · p. 15

B.4 Supplementary clean demonstrations and disjointness

• Supplementary clean demonstrations are recorded only for component-phase cells where the reference policy initially fails, and their configuration determines their split assignment.The post-training buffer retains compiled failure and excluded rows for audit, while positively weighted row counts are reported separately.Source: §B.4 · p. 15, Table 5 · p. 17

B.6 Case-group split

• Each component’s seven collection configurations split into six training and one validation configuration, keeping each case group’s traces together.The resulting split has 60 training and 10 validation case groups per component; validation configurations contain no autonomous-success rows.Source: §B.6 · pp. 15–16

B.5 Post-training buffer composition

• 31/70 GPU rollouts succeeded, compared with 2–11 successes for each of the other components.Table 4 reports the outcome distribution across the 70 autonomous rollouts per component; Table 5 reports compiled training-row counts, not just positively weighted rows.Source: §B.5 · p. 15, Table 4 · p. 16

C.1 Evidence feature definitions

• Process evidence covers partial progress, terminal-region membership, and expected stop events; masks, provenance, safety, and role gates determine row eligibility.Eligible training scores are robustly centered within groups, with empty groups falling back to component-phase and then global eligible training statistics.Source: §C.1 · p. 17, §C.2 · p. 18, §C.3 · p. 18

D.1 Offline validation scoring

• Eight predeclared weighting settings are compared on a fixed planner-phase validation subset; the selected candidate emphasizes terminal-region and stop-signal evidence.Predicted chunks are evaluated from logged start states, and wrong-target or unsafe predictions are blocked from validation credit.Source: §C.5 · p. 18, §D.1 · pp. 18–19, §D.2 · p. 19

D.3 Audits and candidate ranking

• Table 8 ranks candidates by component-balanced Jval on the same fixed 294-row planner-phase validation subset.The pre-training audit checks split leakage, gate consistency, zero imitation weight on gate-failing rows, a finite loss denominator, an upper weight bound, and a non-collapsed effective weighted horizon count.Source: §D.3 · p. 19, Table 8 · p. 20
• All eight candidates pass both audits before their Table 8 ranking is reported.Post-training audits require zero wrong-target and unsafe validation predictions and per-component scoring performance no lower than the reference policy minus a predefined margin.Source: §D.3 · p. 20
• The eight candidates span less than 3 × 10−3 in validation score, and the selected candidate ranks first in 64.7% of resamples.The two leading candidates account for 94.6% of top-ranked outcomes; their score difference is 0.0004 with a 95% interval of [−0.0011, 0.0024].Source: §D.3 · p. 20

E Baseline Specifications

• All baselines start from the shared supervised checkpoint and use the same adapter recipe and hardware protocol.The initial demonstration set trains the shared checkpoint; supplementary clean demonstrations are part of the post-training buffer.Source: E Baseline Specifications · p. 20
• PACER applies its budget to eight predeclared candidates, while each post-training baseline fits one fixed weighting setting.The budget fixes optimizer updates rather than epochs, so smaller-data baselines receive repeated exposure to included rows.Source: E Baseline Specifications · p. 20
• Fixed geometry assigns a precomputed rank-weight scalar to gate-passing rows, independently of PACER coefficients and validation selection.Outcome only uses supplementary clean rows and autonomous successes with uniform row weights.Source: E Baseline Specifications · p. 20
• Selected PACER excludes 227 safety-clamped autonomous rows; uniform all-eligible and fixed geometry include them, yielding 4,386 versus 4,613 positively weighted rows.The selected PACER set has fewer positively weighted training rows than those two baselines.Source: E Baseline Specifications · p. 20

F Training and Implementation

• The policy class, action decoder, and per-step targets remain unchanged; each post-training run trains a new LoRA adapter from the shared π0.5 checkpoint.Fits use a batch size of 32 and a nominal target of 10,000 optimizer updates; control and inference run at 10 Hz.Source: F Training and Implementation · pp. 20–21

G.1 Held-out success with confidence intervals

• PACER matches or exceeds fixed geometry on every component in held-out trials.Figure 4 reports 20 trials per component with 95% Wilson intervals; Connector and CPU show among the largest margins, while the intervals are wide.Source: Figure 4 · p. 22, §G.2 · p. 22

G.2 Per-component comparison

• On CPU fan, PACER reaches ceiling while all baselines remain below it; outcome-only and the no-post-train reference have similar estimates, 11/100 and 12/100.GPU is at ceiling for several methods.Source: §G.2 · p. 22

G.3 Hardware comparison of the eight PACER candidates

• Terminal stop heavy records 74/100 overall and progress heavy 72/100; both reach 30/50 on Setup 2.Their +2 percentage-point difference has a 95% interval of [−6, 11], leaving the leading pair difficult to distinguish overall.Source: §G.3 · p. 22

G.4 Seed and geometry sensitivity

• The three selected-policy seeds achieve 33/50, 35/50, and 32/50, with a mean of 66.7% and standard deviation of 3.1 percentage points.A fit retrained after a 10 mm translation and 0.1 rad rotation reaches 38/50; its difference from seed 42 is +10 points with interval [−7.7, 26.9].Source: §G.4 · pp. 22–23, Table 14 · p. 24

G.5 Annotation consistency

• Two raters agreed on 100, 93, and 98 of three 100-case sets for success, near, and wrong location labels, respectively.The passage says this supports consistency under the labeling rubric.Source: §G.5 · p. 23

G.6 Qualitative rollout sequences

• Figures 5–9 compare five methods across CPU fan, RAM, connector block, GPU, and CPU tasks using endpoint markers for wrong/off-target, near-miss, and target-region outcomes.Figure 5’s sixth row is an unscored illustration of the fixed downstream skill from a reached handoff state.Source: §G.6 · p. 23, Figure 5 · p. 25, Figure 6 · p. 26