Is Online Interaction Necessary for Recovery? A Minimalist Approach to Robust Planning via Perturbation

Bumgeun Park; Donghwan Lee

cs.ROSep 27, 2026

Is Online Interaction Necessary for Recovery? A Minimalist Approach to Robust Planning via Perturbation

Authors: Bumgeun Park, Donghwan Lee

Organizations: Korea Advanced Institute of Science and Technology (KAIST)

Abstract

Behavior cloning (BC) is vulnerable to covariate shift during closed-loop execution, where small prediction or execution errors can drive the robot toward states poorly covered by the demonstration data. We focus on action-sequence planning, where a policy predicts a finite-horizon sequence of actions as a reference trajectory for robot execution. Existing approaches to covariate shift often rely on collecting additional corrective demonstrations, requiring further environment interaction and access to an expert or reference policy. We propose Perturbation-Augmented Recovery Supervision (PARS), a simple training approach for improving recovery using only existing demonstrations. PARS perturbs the robot's proprioceptive state and anchors only the later portion of the predicted trajectory to the original demonstration, leaving the earlier portion free to generate corrective motion and recover toward the demonstrated behavior within the planning horizon. Unlike conventional input-noise augmentation, which preserves the original supervision target over the entire prediction horizon, PARS explicitly provides trajectory-level supervision for recovery from perturbed states. PARS requires neither additional environment interaction nor expert queries and can be instantiated across different action-sequence policy classes with only minor modifications to their BC objectives. Experiments on 51 RLBench manipulation tasks with flow-based, transformer-based, and diffusion-based policies demonstrate that PARS improves robustness to covariate shift during closed-loop execution.

Figures & tables

Figure 1: Standard BC supervises the full predicted trajectory at an unperturbed observation (top). For a perturbed proprioceptive observation, PARS anchors only the predicted suffix to the corresponding demonstrated suffix (bottom). The prefix receives no direct supervision from the perturbed sample, leaving room for corrective motion.

	Method	PARS Comparison		Noise Augmentation	Tasks w/ PARS $\geq$ w/o PARS
	Method	w/o	w/	Noise Augmentation	Tasks w/ PARS $\geq$ w/o PARS
Flow-based	FP	58.0	66.9	60.0	48/51
Flow-based	SFP	49.8	61.6	46.7	44/51
Transformer-based	TP	57.4	67.7	59.3	46/51
Transformer-based	ACT	58.8	68.2	60.0	48/51
Diffusion-based	DP	47.6	56.0	46.2	46/51

Table 1: Performance of PARS across different action-sequence methods on 51 RLBench tasks. Mean success rate (%) over 50 evaluation episodes; for PARS, we report the best per-task augmentation setting. The last column counts the number of tasks on which PARS matches or exceeds the corresponding baseline. Full results for all 51 tasks are provided in Table 8 .

Figure 2: Normalized distance to the demonstrated configurations over the predicted action sequence after deviation. Values below $1.0$ indicate recovery toward the demonstrations, whereas values above $1.0$ indicate further deviation. Curves show the mean over 51 RLBench tasks with $\pm 1$ standard error. PARS promotes recovery or reduces divergence across all evaluated policies.

Figure 3: Kinematics of the free recovery prefix. Distributions of path speed ( $\lVert\Delta q\rVert_{2}/\Delta t$ ) and path acceleration ( $\lVert\Delta^{2}q\rVert_{2}/\Delta t^{2}$ ) over the first $H_{r}=10$ predicted steps. Bars show the fraction of steps in each range, averaged over 51 RLBench tasks. Dashed lines indicate the RLBench planning limits.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Vanilla BC	CCIL	DrilDICE	MBIL	KATS	Stable-BC
0.5	1.2	1.6	0.8	1.7	9.8

Appendix

Table 2: Closed-loop success rates (%) of single-step BC-based baselines, averaged over 51 tasks.

Setting	Value
Training steps	$40{,}000$ gradient steps
Optimizer	AdamW
Learning rate	$3\times 10^{-4}$
Learning-rate schedule	Cosine decay to $10^{-6}$
Batch size	$128$
Random seed	$0$

Appendix

Table 3: Common policy-training and success-rate evaluation settings.

Method	Architecture	Generation / inference	Method-specific settings
FP	Width 144, 4 layers, 8 heads	10 ODE steps	—
SFP	MLP field, width 256	3 Euler substeps per action	$\sigma_{0}=\sigma_{1}=0.1$
TP	Width 152, 4 layers, 8 heads	—	—
ACT	Width 256, 4 layers, 8 heads	—	CVAE latent dim. 32, KL weight 10
DP	U-Net $(128,256,512)$ ; per-task encoder width (256 default)	100 training / 20 inference steps	Squared-cosine noise schedule
KDPE	DP checkpoint	32 candidate chunks	KDE bandwidth 0.2; $20\times$ bandwidth for gripper dimension

Appendix

Table 4: Policy-specific settings. Architecture and policy hyperparameters are selected for the corresponding unaugmented baseline and held fixed across the unaugmented, PARS, and noise-augmentation conditions.

Setting	Applies to	Value
Prediction horizon $H$	All methods	$20$
Free recovery prefix $H_{r}$	PARS	$10$
	Noise augmentation	$0$
Perturbation probability $p$	PARS, noise augmentation	Per task
Perturbation scale range	PARS, noise augmentation	Per task
Inner refinement steps	FP, DP	Per task

Appendix

Table 5: PARS and augmentation settings. Settings marked “per task” are selected separately for each task.

RLBench Task	FP			SFP		TP
RLBench Task	Range	$p$	$K$	Range	$p$	Range	$p$
basketball in hoop	0.05–0.15	0.65	5	0.03–0.08	0.65	0.04–0.12	0.5
beat the buzz	0.02–0.06	0.35	5	0.03–0.08	0.35	0.01–0.04	0.15
change clock	0.03–0.1	0.5	25	0.05–0.08	0.65	0.04–0.12	0.65
close box	0.03–0.1	0.5	25	0.05–0.08	0.65	0.02–0.06	0.5
close door	0.02–0.06	0.5	5	0.05–0.15	0.3	0.05–0.15	0.5

Appendix

Table 6: Task-specific PARS settings for FP, SFP, and TP. Range denotes

\sigma_{\mathrm{lo}}

–

\sigma_{\mathrm{hi}}

p

is the perturbation probability, and

K

is the number of inner refinement steps.

RLBench Task	ACT		DP / KDPE
RLBench Task	Range	$p$	Range	$p$	$K$	Warm-up
basketball in hoop	0.02–0.06	0.35	0.03–0.1	0.5	25	15k
beat the buzz	0.03–0.08	0.1	0.05–0.08	0.5	5	15k
change clock	0.01–0.04	0.25	0.03–0.1	0.5	15	15k
close box	0.02–0.04	0.25	0.03–0.1	0.5	25	15k
close door	0.02–0.04	0.35	0.05–0.15	0.5	15	15k

Appendix

Table 7: Task-specific PARS settings for ACT and DP/KDPE. Range denotes

\sigma_{\mathrm{lo}}

–

\sigma_{\mathrm{hi}}

;

K

is the number of inner refinement steps. Warm-up is reported in gradient steps, with “k” denoting

1{,}000

RLBench Task	Flow-based		Transformer-based		Diffusion-based
RLBench Task	FP	SFP	TP	ACT	DP	KDPE
basketball in hoop	94 $\to$ 96	96 $\to$ 96	94 $\to$ 96	96 $\to$ 96	94 $\to$ 84	96 $\to$ 98
beat the buzz	66 $\to$ 74	48 $\to$ 58	76 $\to$ 72	78 $\to$ 90	28 $\to$ 52	38 $\to$ 50
change clock	6 $\to$ 30	8 $\to$ 22	28 $\to$ 34	18 $\to$ 22	6 $\to$ 24	20 $\to$ 26
close box	88 $\to$ 88	78 $\to$ 90	90 $\to$ 96	92 $\to$ 92	68 $\to$ 88	82 $\to$ 94
close door	12 $\to$ 18	16 $\to$ 16	4 $\to$ 38	12 $\to$ 24	10 $\to$ 14	10 $\to$ 4

Appendix

Table 8: Per-task success rate (%) on all 51 RLBench tasks.

Explore similar work

Sep 24, 2026

RoboRecover: Benchmarking Robot Policy Recovery under Execution Deviations

Robot-policy benchmarks increasingly cover diverse tasks and preset out-of-distribution conditions, but typically evaluate complete trajectories from predefined initial states. These evaluations often focus on the initialized scene and the final outcome, while paying less attention to the dynamic interaction process. During closed-loop execution, actions and contacts can alter object relations and task progress, producing off-nominal intermediate states that need recovery. Recovery requires a policy to infer how task progress has changed, correct the relevant relations, and continue the original goal. We introduce RoboRecover, a benchmark for robot policy recovery under execution deviations. RoboRecover selects deviation states from trajectories, reconstructs them by replaying action prefixes, and evaluates policies on the original task. RoboRecover contains 2,000 scenarios across RoboTwin and LIBERO, with 1,000 scenarios and a fixed 800/200 train/test split on each platform. Results show that initial-state performance does not determine recovery performance and policies exhibit different recovery strengths across scenarios. Using its training split, RoboRecover further supports study on recovery interventions. RoboRecover establishes recovery from execution-induced intermediate states as a distinct dimension of robot policy evaluation.

Yang Li, Chen Zhao, Zhuoran Wang +5

School of Information, Renmin University of China, Beijing, China · Key Laboratory of Data Engineering and Knowledge Engineering, Beijing, China · University of Science and Technology of China +1

Robotic Policy Evaluation Robot Manipulation Benchmarks

Ask PDF

Oct 13, 2026

ReEnvision: Future-Guided Selective Action Repair for Long-Horizon Manipulation

Long-horizon robotic manipulation requires robust execution over sequences of interdependent actions. Existing methods typically handle execution deviations by replanning the remaining action chunk or correcting actions step by step as new observations arrive. However, a local deviation may affect only a few actions under distribution shift, yet replanning regenerates the entire chunk from an unfamiliar state, potentially introducing errors into otherwise valid actions. Step-wise correction may react too late, as it does not explicitly anticipate how the deviation affects subsequent actions. We propose ReEnvision, a future-guided selective action repair framework that updates a joint future-action belief using execution feedback. It then learns how the resulting belief revision affects each unexecuted action. The resulting belief revision enables selective repair of only the affected actions within a cached action chunk, preserving valid actions and reserving full replanning for severe deviations. On CALVIN, L-CALVIN, and RLBench, ReEnvision achieves 86.6% 5-step success, improves 10-step success from 56% to 63%, and reaches an average success rate of 88.8%, respectively. On five real-world manipulation tasks, it achieves an average success rate of 82.8%. Controlled repair experiments further demonstrate the benefits of localizing affected actions and preserving valid ones during execution.

Xiyin Zeng, Xuewei Huang, Yi Lu +2

Information Hub, The Hong Kong University of Science and Technology (Guangzhou) · School of Intelligence Science and Technology, Nanjing University

Long-Horizon Robotic Manipulation Robot Policy Adaptation

Ask PDF

Jul 1, 2026

FAR: Failure-Aware Retry for Test-Time Recovery and Continual Policy Improvement

Robot policies inevitably encounter failures when deployed in real environments. Naive retries often repeat the same mistakes, while many existing recovery methods rely on human intervention. In this paper, we propose Failure-Aware Retry (FAR), a framework that enables robots to learn from previous failures at test time, adapt their behavior accordingly, and eventually complete the task autonomously. FAR combines Failure-Contrastive Preference Adaptation, which constructs preference learning data from failures to steer the policy away from previously unsuccessful behaviors, with lightweight action perturbations during retries to encourage local exploration. We further incorporate successful recovery trajectories into a training loop for continual policy improvement. Experiments in both simulation and real-world manipulation tasks show that FAR substantially improves success rates and robustness, with average gains of 17.6% over the standard diffusion policy in simulation and 11.7% in the real world. In addition, FAR improves data efficiency under both reset and timestep budgets during continual policy improvement by exploiting informative failure cases. Videos and code are available at https://hoar012.github.io/FAR-Project.

Haoran Hao, Shahram Najam Syed, Jeffrey Ichnowski +1

Carnegie Mellon University

Robot Policy Learning Robot Policy Adaptation Robot Skill Learning Robot Failure Recovery Robotic Manipulation Continual RL+1

Ask PDF