ComManip: Overfitting Manipulation Policies to Comfortable Regions

Yidan Lai; Xinyi Chen; Shuquan Man; Huiping Zhuang

cs.ROSep 29, 2026

ComManip: Overfitting Manipulation Policies to Comfortable Regions

Authors: Yidan Lai, Xinyi Chen, Shuquan Man, Huiping Zhuang

Organizations: South China University of Technology, Guangzhou, China · Nanyang Technological University, Singapore

Abstract

Training robot manipulation policies relies on costly robot demonstrations, making large-scale data collection impractical. Meanwhile, to improve policy generalization, existing approaches seek greater diversity in visual observations by varying object placements, viewpoints, and robot configurations during data collection. However, under a limited demonstration budget, this strategy forces the policy to model diverse visual observations, providing insufficient supervision to learn reliable observation-action correspondences under similar local conditions. Our study reveals that policies trained under this strategy achieve lower task success rates than those trained within a compact, visually and kinematically stable region. We refer to these stable regions as comfortable manipulation regions. To exploit this finding, we propose ComManip, a learning paradigm that specializes manipulation policies to comfortable manipulation regions. During inference, ComManip repositions the mobile base until the detected target center enters the familiar image-space range estimated from comfortable-region demonstrations. It then executes the same manipulation policy, enabling effective manipulation across diverse target locations. We conduct extensive experiments across multiple manipulation tasks, demonstration budgets, and policy families including ACT, $π_{0.5}$ , RDT, OpenVLA-OFT, and SmolVLA. The results demonstrate that ComManip improves task success by roughly 20 percentage points or more across different policy architectures in large workspaces under limited demonstration budgets, suggesting that specializing manipulation policies to comfortable regions provides a more data-efficient learning paradigm for manipulation.

Figures & tables

Figure 1: Empirical observation motivating this work. Under the same demonstration budget, concentrating demonstrations within a compact local region yields higher task success rates than distributing demonstrations to cover a larger set of possible object locations.

Figure 2: Overview of ComManip. (1) Demonstrations are collected within a comfortable manipulation region under stable visual and kinematic conditions. (2) A locally specialized policy is trained on these demonstrations. (3) During deployment, the mobile base is repositioned to align the target with the familiar image-space range before policy execution.

Figure 3: Visual alignment for base repositioning. GroundingDINO and SAM2 extract target centers from the initial observations of training data, from which the retained target centers are used to estimate the familiar image-space range $\mathcal{F}_{C}$ and compute the empirical target-center mean $\mu_{C}$ . During deployment, the image-space target center is estimated and compared with $\mathcal{F}_{C}$ . The robot moves the base by a fixed task-specific step in a direction that reduces the image-space deviation from $\mu_{C}$ and re-estimates the target center, repeating this process until it enters $\mathcal{F}_{C}$ .

Figure 4: ACT success rates under matched demonstration budgets for comfortable-region and spatially diverse training and evaluation.

Policy	w ComManip	Stamp Seal	Empty Cup	Object Basket	Phone Stand	Hanging Mug	Stack Blocks	Avg. ( $\Delta_{\mathrm{avg}}$ )
ACT	–	1.6	17.4	5.6	19.0	14.0	8.6	11.0
ACT	✓	69.6	82.4	36.0	62.8	62.8	41.6	59.2 (+48.2)
$\pi$ 0.5	–	48.4	57.8	9.6	42.2	16.8	38.0	35.5
$\pi$ 0.5	✓	65.2	94.4	24.4	54.8	30.8	58.4	54.7 (+19.2)
RDT	–	6.8	29.0	5.8	16.8	19.8	0.6	13.1
RDT	✓	75.2	95.2	28.0	63.2	44.4	48.0	59.0 (+45.9)

Table 1: Task success rates (%) under spatially diverse target placements across the tabletop for six manipulation tasks and five policy families.

\Delta_{\mathrm{avg}}

denotes the average improvement.

Figure 7: Effect of comfortable-region extent on ACT performance. The top row evaluates execution within the comfortable region under large-scale expansion, while the bottom row evaluates execution from the comfortable region to the full workspace under progressive expansion. The corresponding training demonstrations and familiar image-space ranges are shown on the right.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Task	RoboTwin Task Name	Max. Steps
Stamp Seal	stamp_seal	400
Empty Cup	place_empty_cup	500
Object Basket	place_object_basket	700
Phone Stand	place_phone_stand	400
Hanging Mug	hanging_mug	900
Stack Blocks	stack_blocks_two	800

Appendix

Table 4: Official RoboTwin maximum action-step limits for the evaluated tasks.

Task	Skill Type	Object / Target	Main Challenge
Stamp Seal	Planar alignment and contact-rich pressing	Stamp and target region	Accurately aligning the stamp with a small target region before establishing downward contact.
Empty Cup	Grasp-and-place with placement-sensitive release	Cup and coaster	Placing the cup stably on a spatially restricted support region without excessive positional or height error.
Object Basket	Bimanual grasp-and-place	Object and basket	Coordinating object transfer and basket manipulation while keeping the basket upright and avoiding contact with the table.
Phone Stand	Orientation-sensitive constrained placement	Phone and stand	Aligning the phone with the narrow supporting region of the stand before release.
Hanging Mug	Grasping and hook alignment	Mug and rack	Aligning the mug handle with the rack while maintaining sufficient height for successful hanging.
Stack Blocks	Sequential stacking and precision placement	Multiple blocks	Preventing positional errors from accumulating across sequential grasping and stacking stages.

Appendix

Table 5: Evaluated manipulation tasks and their primary technical challenges.

Task	Varied Object	$x$ Range (m)	$y$ Range (m)
Stamp Seal	Stamp	$[-0.15,-0.1],[0.1,0.15]$	$[-0.025,0.025]$
	Target	$[-0.09,-0.04],[0.04,0.09]$	$[-0.025,0.025]$
Empty Cup	Cup	$[-0.2,-0.15],[0.15,0.2]$	$[0,0.05]$
	Coaster	$[-0.02,0.03],[-0.03,0.02]$	$[0,0.05]$
Object Basket	Object	$[-0.25,-0.2],[0.2,0.25]$	$[-0.025,0.025]$
	Basket	$[0.02,0.02]$	$[-0.08,-0.05]$

Appendix

Table 6: Object-placement ranges for the comfortable-region distributions. Positions are expressed in the RoboTwin world coordinate frame. Candidate positions are sampled uniformly within the listed ranges.

Task	Varied Object	$x$ Range (m)	$y$ Range (m)
Stamp Seal	Stamp	$[-0.25,-0.1],[0.1,0.25]$	$[-0.25,0.25]$
	Target	$[-0.19,-0.04],[0.04,0.19]$	$[-0.25,0.25]$
Empty Cup	Cup	$[-0.3,-0.15],[0.15,0.3]$	$[-0.2,0.05]$
	Coaster	$[-0.05,0.1],[-0.1,0.05]$	$[-0.2,0.05]$
Object Basket	Object	$[-0.35,-0.15],[0.15,0.35]$	$[-0.25,0.25]$
	Basket	$[0.02,0.02]$	$[-0.08,-0.02]$

Appendix

Table 7: Object-placement ranges for the workspace-wide distributions. Positions are expressed in the RoboTwin world coordinate frame. Candidate positions are sampled uniformly within the listed ranges.

Task	Varied Object	L1 (m)	L2 (m)	L3 (m)
Empty Cup	Cup	$x:[-0.2,-0.15],[0.15,0.2]$ $y:[0,0.05]$	$x:[-0.25,-0.15],[0.15,0.25]$ $y:[-0.1,0.05]$	$x:[-0.3,-0.15],[0.15,0.3]$ $y:[-0.2,0.05]$
Empty Cup	Coaster	$x:[-0.02,0.03],[-0.03,0.02]$ $y:[0,0.05]$	$x:[-0.02,0.08],[-0.08,0.02]$ $y:[-0.1,0.05]$	$x:[-0.05,0.1],[-0.1,0.05]$ $y:[-0.2,0.05]$
Stamp Seal	Stamp	$x:[-0.15,-0.1],[0.1,0.15]$ $y:[-0.025,0.025]$	$x:[-0.2,-0.1],[0.1,0.2]$ $y:[-0.125,0.125]$	$x:[-0.25,-0.1],[0.1,0.25]$ $y:[-0.25,0.25]$
Stamp Seal	Target	$x:[-0.09,-0.04],[0.04,0.09]$ $y:[-0.025,0.025]$	$x:[-0.14,-0.09],[0.09,0.14]$ $y:[-0.125,0.125]$	$x:[-0.19,-0.04],[0.04,0.19]$ $y:[-0.25,0.25]$
Phone Stand	Phone	$x:[-0.2,-0.15],[0.15,0.2]$ $y:[-0.1,0.05]$	$x:[-0.25,-0.15],[0.15,0.25]$ $y:[-0.2,0.0]$	$x:[-0.25,-0.05],[0.05,0.25]$ $y:[-0.2,0.0]$
Phone Stand	Stand	$x:[-0.05,0],[0,0.05]$ $y:[0.02,0.07]$	$x:[-0.1,0.0],[0,0.1]$ $y:[0.02,0.12]$	$x:[-0.15,0.0],[0,0.15]$ $y:[0,0.2]$

Appendix

Table 8: Physical target-placement ranges used for the large-scale comfortable-region expansion from L1 to L3. Each cell reports the

x

and

y

ranges in the RoboTwin world coordinate frame.

Task	Varied Object	E1 (m)	E2 (m)	E3 (m)	E4 (m)	E5 (m)
Object Basket	Object	$x:[-0.25,-0.23],$ $\phantom{x:}[0.23,0.25]$ $y:[-0.025,0.025]$	$x:[-0.25,-0.22],$ $\phantom{x:}[0.22,0.25]$ $y:[-0.025,0.025]$	$x:[-0.25,-0.2],$ $\phantom{x:}[0.2,0.25]$ $y:[-0.025,0.025]$	$x:[-0.275,-0.22],$ $\phantom{x:}[0.22,0.275]$ $y:[-0.03,0.03]$	$x:[-0.275,-0.22],$ $\phantom{x:}[0.22,0.275]$ $y:[-0.04,0.04]$
Object Basket	Basket	$x:[0.02,0.02]$ $y:[-0.06,-0.05]$	$x:[0.02,0.02]$ $y:[-0.07,-0.05]$	$x:[0.02,0.02]$ $y:[-0.08,-0.05]$	$x:[0.02,0.02]$ $y:[-0.09,-0.05]$	$x:[0.02,0.02]$ $y:[-0.1,-0.05]$
Hanging Mug	Mug	$x:[-0.125,-0.1]$ $y:[-0.01,0.01]$	$x:[-0.15,-0.1]$ $y:[-0.015,0.015]$	$x:[-0.15,-0.1]$ $y:[-0.03,0.03]$	$x:[-0.175,-0.1]$ $y:[-0.03,0.03]$	$x:[-0.175,-0.1]$ $y:[-0.04,0.04]$
Hanging Mug	Rack	$x:[0.1,0.12]$ $y:[0.13,0.14]$	$x:[0.1,0.14]$ $y:[0.13,0.15]$	$x:[0.1,0.15]$ $y:[0.13,0.16]$	$x:[0.1,0.15]$ $y:[0.13,0.17]$	$x:[0.1,0.16]$ $y:[0.13,0.18]$
Stack Blocks	Block 1 Block 2	$x:[-0.2,-0.18],$ $\phantom{x:}[0.18,0.2]$ $y:[-0.04,0.01]$	$x:[-0.2,-0.16],$ $\phantom{x:}[0.16,0.2]$ $y:[-0.04,0.01]$	$x:[-0.2,-0.15],$ $\phantom{x:}[0.15,0.2]$ $y:[-0.04,0.01]$	$x:[-0.23,-0.15],$ $\phantom{x:}[0.15,0.23]$ $y:[-0.04,0.01]$	$x:[-0.25,-0.15],$ $\phantom{x:}[0.15,0.25]$ $y:[-0.04,0.01]$

Appendix

Table 9: Physical target-placement ranges used for the progressive comfortable-region expansion from E1 to E5. Each cell reports the

x

and

y

ranges in the RoboTwin world coordinate frame. All relevant object ranges are varied jointly within each setting.

Panel of Figure 8 — Figure 8: Qualitative localization results across the six evaluated tasks. From left to right, the columns show GroundingDINO detections and confidence scores, pick-target masks, and place-target masks produced by SAM 2.1. The mask centroid is used as the image-space target center. For Stack Blocks, the dash indicates that no separate place-target mask is required during the first pick stage.

Task	Stage / Target	Side	$u$ Range (pixels)	$v$ Range (pixels)
Empty Cup	Cup Pick	L	$[75.5,98.0]$	$[89.0,106.5]$
	Cup Pick	R	$[253.5,273.0]$	$[89.0,107.0]$
	Coaster Place	L	$[160.5,175.0]$	$[112.5,122.0]$
	Coaster Place	R	$[176.0,188.5]$	$[116.0,122.0]$
Phone Stand	Phone Pick	L	$[73.0,101.0]$	$[143.5,161.5]$
	Phone Pick	R	$[250.5,281.5]$	$[142.0,163.0]$

Appendix

Table 10: Task- and stage-specific familiar image-space ranges

\mathcal{F}_{C}=[u_{\min},u_{\max}]\times[v_{\min},v_{\max}]

. L and R denote the left- and right-side configurations, respectively, while Both indicates that the same range is used for both configurations.

Policy	Task	# Demos	Learning Rate	Global Batch	Training Steps	Grad. Accum.
ACT	All six tasks	50 per task	$1\times 10^{-5}$	8	5000	1
$\pi_{0.5}$	Stamp Seal	30	$2.5\times 10^{-5}$	4	5000	1
	Empty Cup	30	$2.5\times 10^{-5}$	8	5000	1
	Object Basket	75	$2.5\times 10^{-5}$	16	5000	1
	Phone Stand	30	$2.5\times 10^{-5}$	8	5000	1
	Hanging Mug	75	$2.5\times 10^{-5}$	16	5000	1

Appendix

Table 11: Task-specific demonstration budgets and training configurations. The final training checkpoint is used for evaluation. For

\pi_{0.5}

, the listed learning rate is the peak learning rate of the training schedule.

Explore similar work

Aug 2, 2026cs.RO

DynamicManip: Enabling Dynamic Manipulation from a Single Static Demonstration

Dynamic manipulation is a critical capability for robots operating in complex and dynamic environments, where robots must interact with objects that are moving or require rapid adjustments. However, learning models for dynamic manipulation tasks face two major challenges: (1) the combinatorial complexity of dynamic scenarios leads to substantial data requirements, and (2) rapid variations in dynamics require real-time and accurate policy execution. In this paper, we propose DynamicManip to address these challenges through an efficient data augmentation pipeline and a low-latency imitation policy. We first propose a static-to-dynamic augmentation pipeline that synthesizes diverse dynamic manipulation demonstrations from a single static demonstration. Second, we introduce a dynamic-aware adaptive policy that adjusts its inference frequency according to task dynamics, enabling responsive and effective dynamic manipulation. Third, we build a dynamic manipulation benchmark, which includes diverse dynamic tasks with an automatic evaluation system for scalable and consistent assessment. Extensive experiments in both simulation and the real world demonstrate that DynamicManip not only provides significant improvements in data efficiency but also achieves better performance in dynamic manipulation tasks, with a mean success rate 18.4 percentage points higher and policy-query latency 32.9% lower.

Haoran Liao, Pengyue Wang, Shuoyu Chen +10

1Sun Yat-sen University · 4Jiangxing Intelligence · 2Stanford University +1

Robotic Manipulation Policies

Ask PDF

Sep 21, 2026cs.RO

Learning Beyond What Humans Can Demonstrate

Behavior cloning for robot manipulation relies on expert demonstrations. However, for tasks that require dynamic stability, precise contact timing, or dexterous coordination, human operators may find it hard or even impossible to collect data. We study this infeasible-demonstration regime and propose GLIDE: Guardrails for Learning from Infeasible Demonstrations Efficiently, a framework that infers task-specific failure modes and converts them into executable guardrails for data collection and policy deployment. Given a task description and the conditioning teleoperation code, GLIDE writes guardrails that use system states to filter teleoperation and policy commands, constrain failure-prone actions, and iteratively improve from trajectory feedback. Across three tasks, GLIDE discovers emergent guardrails that go beyond domain-expert hardcoded ones, improving data collection over naive VR teleoperation and domain-expert hardcoded guardrails. After refinement, GLIDE raises data-collection success from 0-10 percent to 70-90 percent across the three tasks. During policy execution, mixed-data guarded policies reach 70 percent, 60 percent, and 60 percent success on Tomato plate transfer, Marker handover and stand, and Wine serving tasks. These results show that GLIDE can support policy learning when direct demonstrations are infeasible. Project website: http://guardrail-policy.github.io/

Yuchen Song, Aditya Mittal, Unnat Jain

University of California, Irvine

Teleoperation Behavior Cloning Guardrail Latent Failure Patterns

Ask PDF

May 13, 2026cs.RO

SID: Sliding into Distribution for Robust Few-Demonstration Manipulation

Generalizing robotic manipulation across object poses, viewpoints, and dynamic disturbances is difficult, especially with only a few demonstrations. End-to-end visuomotor policies are expressive but data-hungry, while planning and optimization satisfy explicit constraints but do not directly capture the interaction strategies demonstrated by humans. We propose Sliding into Distribution (SID), a structured framework that learns an object-centric motion field from canonicalized demonstrations to iteratively slide the system toward the demonstrated manifold and into the reliable operating region of a lightweight egocentric execution policy, mitigating out-of-distribution (OOD) execution. The motion field provides large corrective motions when far from the demonstration manifold and naturally vanishes near convergence, enabling robust reaching under substantial pose and viewpoint shifts. Within the reached regime, an egocentric policy trained with conditioned flow matching performs task-specific manipulation, supported by kinematically consistent point-cloud reprojection augmentation that preserves action-observation consistency. Across six real-world tasks, SID achieves approximately 90% success under OOD initializations with only two demonstrations, with under a 10% drop under distractors and external disturbances. Overall, SID provides a new paradigm for few-shot manipulation: explicitly managing distribution shift via online distribution recovery.

Yicheng Ma, Wei Yu, Zhian Su +2

1Grasp Lab, Zhejiang University · 2Torch Kernel Co., Ltd.

Robotic Manipulation Policies Egocentric Vision Distributions Manifolds

Ask PDF