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

Chang Liu; Sibo Tian; Litian Gong; Zhiwen Fan; Jiachen Li; Xiao Liang; Minghui Zheng

cs.ROOct 13, 2026

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

Authors: Chang Liu, Sibo Tian, Litian Gong, Zhiwen Fan, Jiachen Li, Xiao Liang, Minghui Zheng

Organizations: Texas A&M University · University of California, Riverside · Georgia Institute of Technology

Abstract

Robotic disassembly of end-of-life (EoL) products requires robots to accommodate product variation, follow prerequisite sequences, and approach components precisely for contact-rich manipulation. Supervised fine-tuning (SFT) provides a practical path for adaptation, but limited and costly disassembly data cannot capture the distribution shifts arising from changing environments and product uncertainties. The policy's own rollouts record its approach attempts, and human corrections demonstrate recovery from its errors, but turning these experiences into better policy performance remains challenging. We introduce PACER, a Process-Aware Correction and Evidence Reweighting framework that learns from these experiences through SFT, using progress toward the target and each trace's source and outcome to weight valid action chunks in the native supervised action loss. PACER trains one policy per weighting setting and selects one of the resulting policies through held-out validation. We evaluate PACER on UR5e motherboard disassembly across five component classes, where it achieves 87/100 successful approach-and-handoff trials on the held-out benchmark, compared with 54/100 for vanilla post-SFT.

Figures & tables

Figure 1: Compared with outcome-only filtering and reward-based adaptation, PACER uses outcome and safety annotations to determine imitation eligibility, then combines execution-process evidence with trace roles to assign bounded weights to valid action chunks in the native supervised action loss.

Figure 2: PACER converts autonomous rollouts, operator corrections, and supplementary clean demonstrations into a weighted native-loss training signal by checking imitation eligibility and weighting valid action chunks with process evidence.

	Overall ( $n=100$ )		Per component ( $n=20$ )					Offline
Method	Success	95% CI (%)	RAM	Conn.	CPU	GPU	CPU fan	$J_{\mathrm{val}}$
Initial SFT $\pi_{\theta_{0}}$	12	$[7.0,19.8]$	7	0	0	5	0	$0.7214$
Vanilla post-SFT	54	$[44.3,63.4]$	12	5	3	20	14	$0.7579$
Uniform all-eligible	64	$[54.2,72.7]$	16	3	8	20	17	$0.7648$
Outcome only	11	$[6.3,18.6]$	0	0	0	10	1	$0.7340$
Clean only	24	$[16.7,33.2]$	0	0	9	8	7	$0.7425$

Table 1: Validation scores and held-out hardware success counts. Each method has 10 trials on each of two configurations per component, one close to the training distribution and one out of distribution. Brackets give marginal Wilson 95% confidence intervals for the overall success rate, in percent. Bold marks success-count column maxima.

	Success counts				PACER gap
Policy	Mild	Strong	Overall	95% CI (%)	$\Delta$ (pp)	95% CI
	$n=50$	$n=50$	$n=100$
PACER selected	43	29	72	$[62.5,79.9]$	Reference	—
Fixed weighting rules
Fixed geometry	35	16	51	$[41.3,60.6]$	$+21$	$[13,29]$
RW-like	33	18	51	$[41.3,60.6]$	$+21$	$[12,30]$

Table 2: Controlled OOD comparison. Each policy has 50 trials on each of two OOD configurations, with mild and strong shifts. Overall intervals are Wilson 95% CIs in percent. Differences are selected PACER minus each row with paired bootstrap 95% CIs. Bold marks success-count column maxima.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Canonical prompt
CPU	Extract the CPU by unlocking the bracket first, then pick up the CPU and place it inside the blue area as a high-value component.
RAM	Extract the closest visible RAM stick and place it inside the blue area as a high-value component.
CPU fan	Extract the CPU fan and place it inside the yellow area as a low-value component.
Connector	Extract the closest visible connector and place it inside the yellow area as a low-value component.
GPU	Extract the graphics card and place it inside the green area, as it needs subsequent disassembly.

Appendix

Table 3: Canonical language prompts by component.

Figure 3: One post-training collection cycle, starting from the initial policy after supervised fine-tuning. A rollout ends in a non-success state (a–c), an operator correction is seeded from that state (d–e), the arm returns to home (f–g), and a supplementary clean demonstration is recorded for the same cell (h–j).

Component	Configs	Rollouts	Success	Partial	Wrong orientation or location	Wrong region or target	Excluded
RAM	7	70	7	3	6	54	0
Connector	7	70	2	25	23	20	0
CPU	7	70	4	8	6	52	0
GPU	7	70	31	37	2	0	0
CPU fan	7	70	11	6	12	41	0

Appendix

Table 4: Per-component outcome distribution of the 70 autonomous rollouts across the seven collection configurations, with the failure count split into its two labels.

Component	Correction rows	Supp. clean rows	All compiled train rows
RAM	301	480	1616
Connector	351	572	2019
CPU	348	587	1880
GPU	107	334	1698
CPU fan	269	364	1623
Total	1376	2337	8836

Appendix

Table 5: Post-training buffer composition as compiled training-row counts after action-chunk compilation. The compiled column counts all rows in the training split, including failure and excluded rows retained for audit, and positively weighted rows per method are reported in Table 9 .

Component	Clean	Correction	Partial	Failure / excluded
RAM	73	46	73	29
Connector	78	73	22	72
CPU	85	45	12	115
GPU	116	69	50	176
CPU fan	45	52	11	86

Appendix

Table 6: Compiled validation rows by component and role. Clean, correction, auto_success, and partial rows form the positive-scoring set

\mathcal{V}_{c}^{+}

, failure or excluded rows are audit-only, and the ranking pass of Table 8 scores a fixed planner-phase subset of these rows.

Candidate	$\beta_{\mathrm{prog}}$	$\beta_{\mathrm{prox}}$	$\beta_{\mathrm{term}}$	$\beta_{\mathrm{stop}}$	$\gamma_{\mathrm{corr}}$	$\gamma_{\mathrm{part}}$	$\tau$	$w_{\max}$
rw_like	0.30	0.45	0.15	0.05	1.20	0.60	1.00	4.50
progress_heavy	0.55	0.20	0.15	0.05	1.20	0.60	1.00	4.50
hover_proximity_heavy	0.20	0.60	0.10	0.05	1.20	0.60	1.00	4.50
terminal_stop_heavy	0.20	0.25	0.35	0.15	1.20	0.60	1.00	4.50
correction_heavy	0.30	0.45	0.15	0.05	1.80	0.70	1.00	4.50
conservative	0.30	0.45	0.15	0.05	1.20	0.40	1.50	2.50

Appendix

Table 7: The eight predeclared weighting settings. The selected candidate emphasizes terminal-region and stop-signal evidence.

Candidate	Internal validation score	Selected
terminal_stop_heavy	0.770760	yes
hover_proximity_heavy	0.770343	no
correction_heavy	0.769622	no
rw_like	0.769558	no
balanced_low_clip	0.769395	no
conservative	0.769244	no

Appendix

Table 8: Candidate ranking by the component-balanced internal validation score

J_{\mathrm{val}}

, computed over the same fixed 294-row planner-phase validation subset for every candidate.

Baseline	Rows used	Train	Horizons	Non-correction rows	Weight	Gate	Selected
$\pi_{\theta_{0}}$ no post-train	none	0	0	0	none	no	no
Vanilla post-SFT	clean, correction	3713	37130	2337	1	no	no
Uniform all-eligible	all eligible	4613	46130	3237	1	yes	no
Outcome only	clean, auto_success	2662	26620	2662	1	yes	no
Clean only	supp. clean	2337	23370	2337	1	yes	no
Correction only	correction	1376	13760	0	1	yes	no

Appendix

Table 9: Baseline specifications. The gate column indicates the use of the PACER eligibility gate

g_{i}

. The train column counts the number of positively weighted training rows, the horizons column sums the mask over those rows, and the non-correction rows column counts positively weighted training rows in the clean, auto_success, or partial roles.

Comparison	Difference	95% CI (pp)
PACER $-$ vanilla post-SFT	33 pp	[14, 49]
PACER $-$ uniform all-eligible	23 pp	[9, 40]
PACER $-$ fixed geometry	8 pp	[3, 12]

Appendix

Table 10: Paired success-rate differences with descriptive component-cluster bootstrap 95 percent intervals over the five components.

Figure 4: Per-component held-out success rate over 20 trials per component. Error bars are 95% Wilson binomial confidence intervals. PACER matches or exceeds fixed geometry on every component.

Candidate	Overall	95% CI (%)	Setup 1	Setup 2
terminal_stop_heavy	$\mathbf{74/100}$	$[64.63,81.60]$	$\mathbf{44/50}$	$\mathbf{30/50}$
progress_heavy	$72/100$	$[62.51,79.86]$	$42/50$	$\mathbf{30/50}$
conservative	$66/100$	$[56.28,74.54]$	$40/50$	$26/50$
ram_connector_recovery	$59/100$	$[49.20,68.13]$	$39/50$	$20/50$
rw_like	$59/100$	$[49.20,68.13]$	$36/50$	$23/50$
hover_proximity_heavy	$55/100$	$[45.24,64.39]$	$36/50$	$19/50$

Appendix

Table 11: Overall and setup-level hardware success for the eight PACER candidates. Each candidate has 100 trials, with 50 per setup. Confidence intervals are marginal Wilson 95% intervals for the overall success rate. Bold marks success-count column maxima.

Candidate	RAM	Connector	CPU	GPU	CPU fan	Overall
terminal_stop_heavy	$\mathbf{15/20}$	$\mathbf{9/20}$	$17/20$	$\mathbf{20/20}$	$13/20$	$\mathbf{74/100}$
progress_heavy	$\mathbf{15/20}$	$6/20$	$14/20$	$18/20$	$\mathbf{19/20}$	$72/100$
conservative	$10/20$	$6/20$	$15/20$	$19/20$	$16/20$	$66/100$
ram_connector_recovery	$12/20$	$4/20$	$15/20$	$13/20$	$15/20$	$59/100$
rw_like	$4/20$	$2/20$	$\mathbf{19/20}$	$16/20$	$18/20$	$59/100$
hover_proximity_heavy	$3/20$	$7/20$	$\mathbf{19/20}$	$12/20$	$14/20$	$55/100$

Appendix

Table 12: Success counts by component pooled across the two hardware setups. Each component has 20 trials per candidate, for 100 trials overall. Bold marks success-count column maxima.

(a) Setup 1
Candidate	RAM	Connector	CPU	GPU	CPU fan	Setup total
terminal_stop_heavy	$5/10$	$\mathbf{9/10}$	$\mathbf{10/10}$	$\mathbf{10/10}$	$\mathbf{10/10}$	$\mathbf{44/50}$
progress_heavy	$\mathbf{7/10}$	$6/10$	$\mathbf{10/10}$	$9/10$	$\mathbf{10/10}$	$42/50$
conservative	$4/10$	$6/10$	$\mathbf{10/10}$	$\mathbf{10/10}$	$\mathbf{10/10}$	$40/50$
ram_connector_recovery	$6/10$	$3/10$	$\mathbf{10/10}$	$\mathbf{10/10}$	$\mathbf{10/10}$	$39/50$
rw_like	$4/10$	$2/10$	$\mathbf{10/10}$	$\mathbf{10/10}$	$\mathbf{10/10}$	$36/50$

Appendix

Table 13: Success counts within each hardware setup. Component cells report successes out of 10 trials; each setup total is out of 50. Bold marks success-count column maxima within each setup.

Training run	Hardware success	95% CI (%)
Selected policy, seed 42	33/50	[52.2,77.6]
Selected policy, seed 43	35/50	[56.2,80.9]
Selected policy, seed 44	32/50	[50.1,75.9]
Geometry perturbation	38/50	[62.6,85.7]

Appendix

Table 14: Seed and geometry sensitivity, with 50 hardware trials per fit and Wilson 95% intervals.

Figure 5: Held-out comparison on the CPU fan task across five methods. The first five rows show the reference policy $\pi_{\theta_{0}}$ , vanilla post-SFT, outcome-only, fixed geometry, and selected PACER, in that order. Red circles mark wrong-location or off-target endpoints, yellow triangles mark near-miss endpoints, and green boxes mark endpoints that reach the target region. The sixth row is a separate, unscored illustration of the fixed downstream skill executed from a reached handoff state.

Figure 6: Held-out comparison on the RAM task across five methods. Row order and endpoint markers are as in Figure 5 .

Figure 7: Held-out comparison on the connector block task across five methods. Row order and endpoint markers are as in Figure 5 .

Figure 8: Held-out comparison on the GPU task across five methods. Row order and endpoint markers are as in Figure 5 .

Figure 9: Held-out comparison on the CPU task across five methods. Row order and endpoint markers are as in Figure 5 .

Robot Policy Learning

Robot policy learning develops controllers that map sensory observations, task goals, and interaction history to actions for embodied robots.

See all 372 papers

ORPA: Online Residual Policy Adaptation for Robot Manipulation Control with Human Feedback
Aug 18, 2026Muhammad A. Muttaqien, Tomohiro Motoda, Ryo Hanai +1Robotic Control Robot Policy Learning
Embodied AI Research Team · National Institute of AIST · Tokyo, Japan
SELF-VLA: A Skill Enhanced Agentic Vision-Language-Action Framework for Contact-Rich Disassembly
Mar 10, 2026Chang Liu, Sibo Tian, Xiao Liang +1Vision-Language-Action Models Contact-Rich Robotic Manipulation
J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA. · Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77843, USA.
RoboDrop: Curating VLA Post-Training Data via Local Gradient Compatibility
Sep 9, 2026Runze Xu, Yuanfan Xu, Cuijie Xu +4Robotic Data Collection VLA Model Post-Training
Tsinghua University · Striding AI

Aug 18, 2026J/K move · Enter open · S save

ORPA: Online Residual Policy Adaptation for Robot Manipulation Control with Human Feedback

Muhammad A. Muttaqien, Tomohiro Motoda, Ryo Hanai, Yukiyasu Domae

Embodied AI Research Team · National Institute of AIST · Tokyo, Japan

Robotic Control Robot Policy Learning Robot Policy Adaptation Robotic Manipulation Residual Policy Learning

Open paper AskPDF arXiv

Robotic manipulation policies trained via imitation learning, such as Action Chunking with Transformers (ACT), can achieve strong performance under ideal conditions but often remain sensitive to small execution errors and distribution shifts. Correcting these failures typically requires dataset aggregation and full-policy retraining, which is computationally expensive and unsuitable for real-time deployment. In this work, we propose Online Residual Policy Adaptation (ORPA), a framework that enables immediate, feedback-driven correction of robot actions without modifying the underlying policy parameters. ORPA augments a pretrained control policy with a lightweight, feedback-conditioned module that predicts residual adjustments directly in joint space, allowing the system to adapt its behavior at runtime. We evaluate ORPA on a set of precision-sensitive manipulation tasks using the ALOHA platform, demonstrating improvements in success rate and recovery from small perturbations compared to baseline control policies and rule-based inverse kinematics corrections.

Abstract

Robot Policy Learning

Explore similar work

ORPA: Online Residual Policy Adaptation for Robot Manipulation Control with Human Feedback

SELF-VLA: A Skill Enhanced Agentic Vision-Language-Action Framework for Contact-Rich Disassembly

RoboDrop: Curating VLA Post-Training Data via Local Gradient Compatibility