PaperlayerView exact revision on arXiv

Source-linked AI summary

ComManip: Overfitting Manipulation Policies to Comfortable Regions

Yidan Lai, Xinyi Chen, Shuquan Man, Huiping Zhuang

arXiv:2609.36928v1cs.RO

TL;DR

Robot demonstrations are costly, while broad target-placement coverage can thin supervision under limited budgets. ComManip trains locally specialized policies in comfortable regions and repositions the base to bring targets into the familiar image-space range. Across five policy families, it reports an average gain of 40.9 percentage points under matched demonstration budgets.

  • Problem

    With limited demonstrations, covering more target locations spreads examples across visual observations and trajectories, leaving fewer examples under similar local conditions.

    Source: §1 · p. 2

  • Method

    ComManip trains a locally specialized policy on comfortable-region demonstrations and repositions the mobile base to align targets with the familiar image-space range before executing it.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    ComManip yields an average gain of 40.9 percentage points across five policy families under matched demonstration budgets.

    Source: §4.4 · p. 8

  • Takeaways & Limitations

    The results support combining local policy learning with base repositioning to handle target-location variation.

    Source: §5 · p. 10

Abstract

from arXiv · show

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, $\pi_{0.5}$, 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.

1 INTRODUCTION

Robot demonstrations are costly, and broadening target-placement coverage can thin supervision under limited budgets. The paper finds that compact-region demonstrations yield higher success, motivating ComManip’s locally specialized policies and deployment-time base repositioning.

1 INTRODUCTION

  • Robot demonstrations require task execution, environment resets, and quality control, making data-efficient learning under limited budgets important.General-purpose policies also require datasets that cover varied operating conditions.Source: §1 · p. 1
  • Limited demonstration budgets spread examples across more visual observations and trajectories, leaving fewer demonstrations under similar local conditions.The paper questions whether spatial diversity remains beneficial when demonstrations are limited.Source: §1 · p. 2
  • Under the same demonstration budget, concentrating demonstrations in a compact region yields higher task success than covering more object locations.The authors call compact physical-workspace regions where limited demonstrations support reliable learning comfortable manipulation regions.Source: §1 · p. 2, Figure 1 · p. 2
  • ComManip trains locally specialized policies within comfortable regions, then repositions the mobile base to align out-of-region targets with the familiar image-space range.The same manipulation policy is executed after alignment.Source: §1 · p. 2
  • The introduction reports experiments across tasks, demonstration budgets, and ACT, π0.5, RDT, OpenVLA-OFT, and SmolVLA, with consistent improvements under diverse target placements.The contribution summary characterizes direct learning from spatially diverse placements as less data-efficient under limited budgets.Source: §1 · p. 2

2 RELATED WORK

Prior work expands manipulation learning through larger, more diverse datasets and policy architectures, and addresses generalization through data collection, representations, or base-pose selection. ComManip instead combines comfortable-region specialization with deployment-time repositioning.

2 RELATED WORK

  • Figure 2 summarizes ComManip as comfortable-region demonstration collection, local policy learning, and deployment-time base repositioning to align the target with the familiar image-space range.The region is selected to maintain favorable visual and kinematic conditions.Source: Figure 2 · p. 3
  • Large-scale datasets and generalist or VLA policies use diverse data, language conditioning, pretraining, or expressive action modeling; ACT and Diffusion Policy use action chunking and generative trajectory modeling.These lines of work advance manipulation learning through larger datasets and policy architectures.Source: §2 · pp. 2–3
  • ComManip concentrates demonstrations in a compact region and uses deployment-time base repositioning to address workspace variation without changing the policy architecture or learning objective.This contrasts with cited approaches that improve generalization through data collection or representation learning.Source: §2 · p. 3
  • Prior mobile-base methods select or optimize poses using scene reconstruction, spatial representations, affordances, reachability, or learned pose preferences.Mobi-π repositions a robot to recover observations compatible with a pretrained policy’s training distribution.Source: §2 · p. 3

3 METHOD

ComManip trains a locally specialized policy from demonstrations in a comfortable region, then uses deployment-time visual alignment to handle targets outside that region. Its method defines the region and training-image statistics, allocates the demonstration budget locally, and repositions the base until the target enters a familiar image-space range.

3.1 FRAMEWORK OVERVIEW

  • ComManip aligns an out-of-region target with the familiar image-space range, then executes the same locally trained policy without modification.The framework combines comfortable-region demonstration collection, local policy learning, and deployment-time base repositioning.Source: §3.1 · p. 4, §3.4 · p. 6

3.2 COMFORTABLE MANIPULATION REGION

  • A comfortable manipulation region is a compact workspace area with clear target visibility and favorable kinematics, where target placements are sampled for successful demonstrations.The same demonstrations train the policy and estimate deployment target-center statistics, so no separate calibration set is needed.Source: §3.2 · p. 4

3.3 COMFORTABLE-REGION-CENTRIC LEARNING

  • ComManip allocates the fixed budget of successful demonstrations to the selected region and trains the policy with its original learning objective.This favors local specialization, while workspace variation is handled at deployment through base repositioning.Source: §3.3 · p. 5

3.4 MOBILE BASE REPOSITIONING

  • At deployment, the robot moves the base by fixed task-specific steps to reduce the target-center deviation from the training mean until the center enters the familiar range or Kmax movements are reached.After alignment, the base stops and the policy executes unchanged; the alignment rule does not estimate an optimal base pose and is compatible with different policy architectures.Source: §3.4 · pp. 5–6, Figure 3 · p. 5

4 EXPERIMENTS

The experiments test whether comfortable-region specialization remains effective across tasks, policy families, deployment settings, region extents, and demonstration budgets. ComManip consistently improves results under spatially diverse target placements, while the region-extent and budget studies show that its performance depends on task and expansion scale.

4.1 EXPERIMENTAL SETUP; 4.2 EVALUATION PROTOCOL AND COMPARED SETTINGS

  • Evaluation covers six simulation tasks and two real-world tasks, using ACT, π0.5, RDT, OpenVLA-OFT, and SmolVLA; baseline and ComManip comparisons use matched budgets and training settings.The main policy comparison reports mean success across five random-seed runs, with 100 episodes per run.Source: §4.1 · pp. 6–7, §4.2 · p. 7

4.4 MAIN RESULT

  • Across five policy families, ComManip yields a 40.9 percentage-point average gain; task-averaged gains range from 25.8 to 64.3 pp.Table 1 compares task success rates under spatially diverse placements with matched demonstration budgets.Source: Table 1 · p. 7, §4.4 · p. 8

4.3 DATA EFFICIENCY UNDER SPATIAL DIVERSITY

  • Under matched budgets, comfortable-region training outperforms spatially diverse training on all three tasks; 15 Object Basket demonstrations match spatially diverse training with 5× as many.Figure 4 compares ACT success rates for comfortable-region and spatially diverse training and evaluation.Source: §4.3 · p. 7, Figure 4 · p. 7

4.5 EFFECT OF REPOSITIONING

  • For ACT, ComManip raises average local success from 77.2% to 86.7% and achieves 71.6% across diverse placements versus 12.7% for direct training.Real-world success also rises from 10.0% to 73.3% on Bottle Placing and from 6.7% to 56.7% on Blocks Stacking.Source: §4.5 · p. 8, Table 2 · p. 8, Table 3 · p. 9

4.6 EFFECT OF COMFORTABLE-REGION EXTENT

  • With large-scale expansion, performance generally declines as the region grows, and both approaches degrade at L3; with local expansion, ComManip leads across extents, while Hanging Mug peaks at E3.The results suggest balancing demonstration concentration with sufficient local variation.Source: §4.6 · pp. 8–9, Figure 7 · p. 9

4.7 EFFECT OF DEMONSTRATION QUANTITY

  • Across evaluated tasks and budgets, ComManip matches or exceeds spatially diverse training; more comfortable-region demonstrations generally help, but higher-budget trends vary by task.On Object Basket, increasing the budget from 15 to 75 raises the success rate to 2× its value at the initial budget.Source: Figure 5 · p. 8, §4.7 · p. 9

5 CONCLUSION

Under limited demonstration budgets, the paper finds locally specialized policies more data-efficient than policies trained across spatially diverse target placements. ComManip pairs comfortable-region policy learning with base repositioning and shows consistent improvements across tasks and policy families.

5 CONCLUSION

  • Under limited budgets, policies trained on spatially diverse target placements are less data-efficient than locally specialized policies trained in stable regions.The conclusion supports combining local policy learning with base repositioning to handle target-location variation.Source: §5 · pp. 9–10
  • ComManip combines comfortable-region policy learning with base repositioning to recover familiar visual and kinematic conditions at deployment.The experiments report consistent improvements under matched demonstration budgets across six tasks and five policy families.Source: §5 · p. 10
  • Further analyses examine the effects of repositioning, comfortable-region extent, and demonstration quantity.Source: §5 · p. 10

A EXPERIMENTAL SETUP

The experiments use RoboTwin 2.0 with a fixed-base AgileX ALOHA embodiment and the SAPIEN physics backend. The setup specifies the camera configuration and reports official task action-step limits in Table 4.

A EXPERIMENTAL SETUP

  • Experiments use RoboTwin 2.0, the fixed-base AgileX ALOHA embodiment, and the SAPIEN physics backend.The robot has two 6-DoF manipulators with parallel grippers.Source: §A.1 · p. 13
  • The default D435 head camera provides 320 × 240 RGB observations with a 37° vertical field of view.The passage also specifies its initialization, direction vectors, and clipping planes.Source: §A.1 · p. 13
  • Table 4 summarizes the official RoboTwin maximum action-step limits for the evaluated tasks.Source: Table 4 · p. 14

A.2 BASE-REPOSITIONING IMPLEMENTATION

Because the RoboTwin ALOHA embodiment is fixed-base, the implementation emulates repositioning by translating the robot and camera assembly within the tabletop plane.

A.2 BASE-REPOSITIONING IMPLEMENTATION

  • Base repositioning is emulated by translating the fixed-base robot and camera assembly together while preserving their relative transforms and camera orientation.Source: §A.2 · p. 14
  • Each movement uses a task- and stage-specific 0.002 or 0.005 m step interpolated over eight simulation frames.When initial object normalization is enabled, two additional settling steps are executed.Source: §A.2 · p. 14

A.3 EVALUATION PROTOCOL

Evaluation follows the official RoboTwin protocol, with task-specific success checks and action-step limits. Stamp Seal additionally specifies a planar position tolerance and the positions used in its criterion.

A.3 EVALUATION PROTOCOL

  • Success is determined by each task’s success() check and applied consistently to policies, baselines, and ComManip variants.Source: §A.3 · p. 14
  • For Stamp Seal, the protocol specifies a 0.02 m planar position tolerance and defines stamp and target positions in world coordinates.Source: §A.3 · p. 14
  • Unsuccessful episodes terminate at the task-specific maximum action-step limit, with the official limits listed in Table 4.Success evaluation is automatic.Source: §A.3 · p. 14

A.4 MANIPULATION TASKS

The appendix describes the evaluated tasks and policy families, then details how successful demonstrations and target placements are configured. It also specifies the physical settings used to expand comfortable regions.

A.4 MANIPULATION TASKS

  • The six evaluated RoboTwin tasks are summarized by their primary manipulation challenges, but those challenges are not enumerated here.Table 5 provides the task-and-challenge summary.Source: §A.4 · p. 14, Table 5 · p. 15

A.5 POLICIES

  • ACT, π0.5, RDT, and OpenVLA-OFT use RoboTwin 2.0 implementations, while SmolVLA uses its released base checkpoint with recommended fine-tuning.All policies receive three RGB streams and use the same 14-dimensional joint-position action interface, while retaining model-specific preprocessing and representations.Source: §A.5 · p. 14

A.6 DEMONSTRATION COLLECTION

  • Demonstrations come from RoboTwin’s automated expert framework; only successful rollouts are retained in pools shared across policy families and matched-budget comparisons.Comfortable-region and workspace-wide variants use matched counts of successful demonstrations.Source: §A.6 · pp. 14–15

A.7 TARGET-PLACEMENT DISTRIBUTIONS

  • Object positions are sampled uniformly within task-specific bounds and filtered through RoboTwin validity checks; orientations follow official task configurations.Tables 6 and 7 specify comfortable-region and workspace-wide placement ranges.Source: §A.7 · p. 15, Table 6 · p. 15, Table 7 · p. 16

A.8 COMFORTABLE-REGION EXTENT SETTINGS

  • The extent study uses L1–L3 for large-scale expansion on simple tasks and E1–E5 for progressive expansion on complex tasks.For multi-object tasks, all listed object ranges vary jointly within each setting.Source: §A.8 · p. 16, Table 8 · p. 16

B IMPLEMENTATION DETAILS

Implementation details cover target localization, image-to-world alignment, deployment ranges, failure handling, and policy-state resets. Localization uses semantic detections refined into mask-based target centers.

B.1 OBJECT LOCALIZATION AND SEGMENTATION

  • GroundingDINO-SwinT selects the highest-confidence retained target box after both box and text scores exceed 0.25; SAM 2.1 refines it into a mask whose centroid defines the image-space target center.The procedure uses task- and stage-specific prompts and is illustrated qualitatively across six tasks.Source: §B.1 · p. 16, Figure 8 · p. 18

B.2 IMAGE-TO-WORLD MOVEMENT MAPPING

  • Image-space corrections map to planar robot translations using task- and stage-specific step sizes of 0.002 or 0.005 m and bounds of the familiar range.When both coordinates fall outside the range, both translations apply in one iteration; each movement is followed by a fresh observation and center estimate.Source: §B.2 · p. 17

B.3 TASK-SPECIFIC FAMILIAR RANGES

  • Deployment uses task- and stage-specific familiar image-space ranges, with separate ranges for left/right configurations when applicable.Coordinates are in pixels in the original 320 × 240 head-camera image.Source: §B.3 · p. 17

B.4 FAILURE HANDLING

  • Alignment fails after three consecutive detection misses or when the target has not entered the familiar range within Kmax movements.The miss limit applies after at least one successful detection.Source: §B.4 · p. 17

B.5 POLICY-STATE RESET

  • After repositioning, policies are reset to prevent alignment observations or action history from carrying over, then receive a fresh camera observation.ACT, SmolVLA, RDT, and π0.5 have their respective temporal, queue, or observation-window state cleared; OpenVLA-OFT requires no reset.Source: §B.5 · p. 17

C ADDITIONAL TRAINING DETAILS

The appendix records policy training configurations and computational resources. It retains each policy family’s original training procedure, with ACT trained from scratch and the remaining policies fine-tuned from released checkpoints.

C ADDITIONAL TRAINING DETAILS

  • ACT is trained from scratch with an ImageNet-pretrained ResNet-18 visual backbone, while the other policies are fine-tuned from released base checkpoints.Table 11 summarizes task-specific budgets, learning rates, batch sizes, training lengths, gradient accumulation, and evaluation checkpoints.Source: C ADDITIONAL TRAINING DETAILS · pp. 18–19, Table 11 · p. 20

C.1 COMPUTATIONAL RESOURCES

  • Training uses one NVIDIA H100 GPU with 80 GB memory, and policy evaluation uses one NVIDIA GeForce RTX 3090 GPU.Reported batch sizes are global batch sizes after gradient accumulation.Source: §C.1 · p. 20

Related papers

Loading 2609.36928v1…