PaperlayerView exact revision on arXiv

Source-linked AI summary

Imperfection for Precision: Upcycling Imperfect Data for High-Precision Robotic Manipulation

Hao Wei, Yang Liu, Chao Tang, Shengbao Li, Jiangtao Chen, Jinxuan Zhu, Jiaheng Wang, Hong Yin, Zhaofeng Cao, Tingguang Li

arXiv:2609.26672v1cs.RO

TL;DR

High-precision manipulation relies on costly task-specific demonstrations, while imperfect data can be difficult to reuse directly. ε4P assigns different kinds of imperfect data to source-dependent regions of the flow-matching trajectory. Real-robot experiments report performance gains and show that some high-quality data can be replaced with a small average drop.

  • Problem

    High-precision manipulation requires costly task-specific, high-quality demonstrations, while available imperfect data often falls short of the precision required for direct reuse.

    Source: §I · p. 1

  • Method

    ε4P routes low-precision target-task data and high-precision task-mismatched data to different flow-time regions for training a policy.

    Source: §I · p. 1, B. How Flow Time Shifts Action-Context… · p. 3

  • Results

    Real-robot experiments show gains of up to 31.7 percentage points over training with perfect target-task data alone and an average drop of 4.2 percentage points when replacing an equal amount of additional perfect data.

    Source: §I · p. 2, §VII · p. 7

  • Takeaways & Limitations

    The findings support repurposing heterogeneous imperfect data to reduce reliance on costly task-specific, high-quality data for high-precision manipulation.

    Source: §I · p. 2

    Experiments use a single robot platform and relatively few demonstrations in a controlled multicamera setup, and do not fully evaluate scalability or cross-embodiment generality.

    Source: §VI · p. 7

Abstract

from arXiv · show

Training vision-language-action (VLA) models for high-precision manipulation typically requires task-specific, high-quality data (e.g., teleoperation), which is slow and expensive to collect. To reduce this burden without compromising manipulation precision, we propose $\varepsilon$4P (Imperfection for Precision), a simple yet effective method that "upcycles" two otherwise discarded data sources: (1) low-precision data from the target task and (2) high-precision data from mismatched tasks. Rather than naively mixing these imperfect data sources throughout co-training, $\varepsilon$4P controls where each source contributes along the flow-matching trajectory. Specifically, low-precision, target-task data is used at high noise to preserve high-level task context and high-precision, task-mismatched data is used at low noise to transfer low-level action precision. Through real-robot experiments on both sub-millimeter, high-precision tasks and coarse-grained tasks, we demonstrate that the proposed method (1) effectively leverages additional imperfect data to improve policy performance by up to 31.7 percentage points, and (2) can replace an equal amount of task-specific, high-quality data with an average performance drop of only 4.2 percentage points. Overall, $\varepsilon$4P points toward a scalable paradigm for high-precision manipulation, in which heterogeneous, imperfect data can be systematically repurposed to reduce reliance on costly task-specific, high-quality data. More details are available at https://varepsilon4p.github.io/.

I. INTRODUCTION

High-precision manipulation depends on costly task-specific demonstrations, while abundant imperfect data is difficult to reuse because it can conflict with precise behavior. ε4P routes two kinds of imperfect data to different regions of the flow-matching trajectory; real-robot experiments show improved performance and replacement of some high-quality data with a small average drop.

I. INTRODUCTION

  • Task-specific, high-quality demonstrations are costly to collect, while imperfect data often falls short of the precision required for direct reuse.Uniformly mixing imperfect sources can create conflicting supervision and bias policies away from precise behaviors.Source: §I · p. 1
  • ε4P routes low-precision target-task data and high-precision task-mismatched data to source-dependent admission regions along the flow-matching trajectory.It identifies where noised target-task actions become sufficiently indistinguishable from high-precision demonstrations, and uses the transition from context-dominant to action-dominant prediction for mismatched-task data.Source: §I · pp. 1–2, Figure 1 · p. 1, Figure 2 · p. 2
  • The method integrates heterogeneous imperfect data into one policy while limiting their imperfections, without changing the model architecture or flow-matching objective.Its intended role is to reduce reliance on costly task-specific, high-quality data.Source: §I · p. 2
  • ε4P improves performance by up to 31.7 percentage points and replaces an equal amount of task-specific, high-quality data with an average drop of 4.2 percentage points.Real-robot experiments cover sub-millimeter cable plugging and connector insertion, as well as coarse-grained Bolt-Nut Sorting.Source: §I · p. 2

II. RELATED WORK

Prior work controls how much heterogeneous or imperfect data contributes, generally applying its contribution uniformly across the training objective. ε4P instead schedules each imperfect source by flow time, building on noise-dependent scheduling while providing a mechanism to derive source-specific admission regions.

A. Heterogeneous Data Co-Training

  • Prior co-training approaches reweight domains or select samples and demonstrations, but generally apply each source’s contribution uniformly across the training objective.Uniform integration can introduce conflicting supervision and bias a policy away from behaviors needed for high-precision manipulation.Source: A. Heterogeneous Data Co-Training · p. 2

B. Noise and Flow-Time Scheduling

  • Diffusion and flow-time research shows that the utility of information and the influence of corruption can vary across noise levels, including in sequential decision making.Prior robotics approaches use different noise levels across tokens, action horizons, or conditioning strength.Source: B. Noise and Flow-Time Scheduling · p. 2
  • ε4P differs from prior noise-scheduling approaches by automatically deriving source-dependent admission regions for precision and task mismatch.The related work contrasts this mechanism with Ambient Diffusion Policy, which also uses diffusion time to leverage suboptimal data.Source: B. Noise and Flow-Time Scheduling · p. 3

III. APPROACH

The approach trains a flow-matching policy with high-precision target-task data and two imperfect sources, assigning each an admission region along flow time. Its design draws on the observed shift from action reliance at low flow times to context reliance at high flow times.

A. Problem Formulation

  • The training setup combines high-precision target-task data with precision-mismatched target-task data and high-precision task-mismatched data.The two imperfect sources are instantiated with UMI and teleoperation data, respectively.Source: A. Problem Formulation · p. 3
  • Each source has an admission region, with target-task high-precision data admitted across [0, 1]; native co-training admits every source across that full interval.The objective is to choose regions for the imperfect sources so each contributes only where it provides useful supervision.Source: A. Problem Formulation · p. 3

B. How Flow Time Shifts Action-Context Reliance

  • The reliance analysis perturbs the noised action input or task context and measures changes in the predicted action across flow time.The final reliance scores average the respective sensitivities and normalize them by sensitivity when both inputs are perturbed.Source: B. How Flow Time Shifts Action-Context… · p. 3
  • The policy’s reliance shifts from action-dominant prediction at low flow times to context-dominant prediction at high flow times, with one crossover.This suggests low-flow supervision from task-mismatched, high-precision data and high-flow supervision from target-aligned, lower-precision data.Source: B. How Flow Time Shifts Action-Context… · p. 3, Figure 3 · p. 3

C. Offline Admission Boundaries

ε4P derives separate admission boundaries for the two imperfect sources because task mismatch and precision mismatch require different diagnostics. It admits task-mismatched data at low flow times and precision-mismatched data at high flow times.

C. Offline Admission Boundaries

  • The two sources require different diagnostics: action separability for precision-mismatched Dpm and action-context reliance for task-mismatched Dtm.The distinction reflects the different ways each source differs from target-task data.Source: C. Offline Admission Boundaries · pp. 3–4
  • For Dpm, a flow-time-conditioned classifier compares noised actions with Dpt; tpm is the earliest time its held-out AUC is within σ = 0.05 of chance.The resulting admission region is Tpm = [tpm, 1], retaining Dpm at high flow times.Source: C. Offline Admission Boundaries · pp. 3–4
  • For Dtm, ttm is the latest flow time when action reliance remains at least as strong as context reliance, defining the low-flow admission region Ttm = [0, ttm].At low flow times, prediction relies primarily on noised action input; context becomes more influential as flow time increases.Source: C. Offline Admission Boundaries · p. 4

D. Policy Integration

ε4P samples each training example’s flow time from an admission region determined by its source, then uses the standard flow-matching objective. The target-task source spans the full trajectory, while imperfect sources are restricted to their designated regions.

D. Policy Integration

  • At each step, ε4P samples a source and example, then draws flow time uniformly from that source’s admission region.The regions are [0, 1] for Dpt, [tpm, 1] for Dpm, and [0, ttm] for Dtm.Source: D. Policy Integration · p. 4
  • The sampled flow time constructs xt, after which the policy is optimized with the standard flow-matching objective.Dpt supervises the full trajectory, while each imperfect source contributes only within its admitted region.Source: D. Policy Integration · p. 4
  • Three-Prong Plug Insertion supplies shared auxiliary dataset Dtm across the ATX, Cable, and Sorting target tasks.The target tasks are ATX 24-Pin Connector Insertion, Two-Stage Cable Plugging, and Bolt-Nut Sorting.Source: Figure 4 · p. 4

IV. EXPERIMENTAL SETUP

The real-robot evaluation spans two sub-millimeter insertion tasks and a coarse-grained sorting task, using teleoperation and UMI demonstrations. Dataset composition and native co-training success rates are presented in separate tables.

IV. EXPERIMENTAL SETUP

  • Experiments use an RB-Y1 robot, with high-precision demonstrations collected by teleoperation and lower-precision demonstrations collected using a self-built UMI system.Both data sources use visual observations from two Intel RealSense D405 cameras; UMI camera placement and gripper configuration match the robot’s.Source: §IV · p. 4
  • ATX insertion has approximately 0.2 mm nominal clearance, and the second stage of Cable plugging has approximately 0.3 mm lateral clearance.Bolt-Nut Sorting has substantially larger placement tolerance and serves as a coarse-grained benchmark.Source: §IV · p. 4
  • UMI provides precision-mismatched Dpm for each target task, while three-prong plug insertion provides shared task-mismatched Dtm.Dtm is used as auxiliary data across the target tasks.Source: §IV · p. 4
  • Table I covers real-robot training-dataset composition, while Table II covers success rates under native co-training.The supplied captions identify each table’s scope without providing its cell values.Source: Table I · p. 5, Table II · p. 5

V. EXPERIMENTS

The experiments are organized around four questions about imperfect data: its effects under native co-training, ε4P’s ability to improve performance, the admission regions, and replacement of high-quality data.

V. EXPERIMENTS

  • The first question asks whether imperfect data helps or harms policy learning under native co-training.The setup defines native co-training as admitting each source across the full flow-time axis.Source: §V · p. 5
  • The second question asks whether ε4P can upcycle imperfect data to improve policy performance.Source: §V · p. 5
  • The third question asks whether the identified admission regions capture where imperfect data is useful.Source: §V · p. 5
  • The fourth question asks whether ε4P can replace some high-quality data with imperfect data.Source: §V · p. 5

A. Imperfect Data Is Harmful under Native Co-Training

Native co-training can degrade performance, especially with precision-mismatched data; ε4P’s source-specific admission turns imperfect-data supervision into gains, with both sources together performing strongly.

B. ε4P Upcycles Imperfect Data to Improve Performance

  • 80.0% success on ATX, 91.7% on Sorting, and 88.3% full-task success on Cable; the combined policy outperforms native co-training and Dpt alone.On ATX, composition also outperforms either imperfect source alone; on Sorting and Cable, it retains nearly all single-source gains.Source: Table III · p. 5, B. ε4P Upcycles Imperfect Data to… · p. 6
  • 73.3% ATX success with Dpm admitted only in Tpm, up from 33.3%; Dtm admission in Ttm raises ATX success from 43.3% to 71.7%.The gains extend to the other tasks, and Dpm’s relative performance reverses: it is more harmful than Dtm under native co-training but performs better across tasks under ε4P.Source: Table III · p. 5, B. ε4P Upcycles Imperfect Data to… · p. 5

A. Imperfect Data Is Harmful under Native Co-Training

  • Native co-training with Dpm reduced success by 34.2 percentage points on average on ATX and Cable versus 6.7 on Sorting; Dtm drops were 5.0, 3.4, and 6.6 points.All sources were admitted over the full flow-time axis; degradation therefore varied with both data mismatch and task precision requirements.Source: A. Imperfect Data Is Harmful under… · p. 5

B. ε4P Upcycles Imperfect Data to Improve Performance

  • Reducing Dpm’s global weight mitigates native co-training’s degradation, but no nonzero weight beats Dpt alone; ε4P reaches 73.3% on ATX by restricting where Dpm contributes.The comparison keeps Dpm at full weight under ε4P, isolating flow-time support rather than contribution amount.Source: Figure 5 · p. 6, B. ε4P Upcycles Imperfect Data to… · p. 6

C. ε4P Identifies Effective Admission Regions

Dense sweeps, a local boundary perturbation, and opposite-window controls validate that ε4P’s source-dependent admission regions align with effective downstream policy performance.

C. ε4P Identifies Effective Admission Regions

  • On ATX, both identified boundaries, tpm and ttm, lie within the high-performing regions in dense sweeps.The downstream sweeps validate boundaries computed using offline diagnostics.Source: C. ε4P Identifies Effective Admission… · p. 6
  • On Cable, the identified ttm = 0.4 achieves the highest success among the evaluated values, 0.2, 0.4, and 0.6.This provides local validation on a second task.Source: C. ε4P Identifies Effective Admission… · p. 6
  • On ATX, the selected windows outperform opposite windows and native co-training: Tpm yields 73.3% versus 30.0% and 33.3%, while Ttm yields 71.7% versus 46.7% and 43.3%.The results indicate that performance depends on where each source is admitted, not merely on truncating its flow-time range.Source: C. ε4P Identifies Effective Admission… · p. 6

D. ε4P Reduces Reliance on High-Quality Data

ε4P’s source-dependent admission can substitute for an equal amount of additional high-quality data with a small average success-rate drop, though replacement performance depends on mismatch type.

D. ε4P Reduces Reliance on High-Quality Data

  • 4.2 percentage points average success-rate drop across both imperfect-data sources and all three tasks when replacing an equal amount of additional high-quality data with admitted imperfect data.Dpm recovers nearly all the benefit of the additional perfect data; Dtm’s gaps are 3.3 points on ATX, 6.7 on Sorting, and 10.0 on full Cable.Source: Table V · p. 7, D. ε4P Reduces Reliance on… · p. 7

VI. DISCUSSION AND FUTURE WORK

The discussion argues that imperfect data are useful only in source-appropriate flow-time regions, while the conclusion reports gains from this strategy and modest performance loss when replacing perfect data. The experiments remain limited in scale and embodiment.

VI. DISCUSSION AND FUTURE WORK

  • Imperfect auxiliary data provide useful or harmful supervision at different flow times, and opposite-window and loss-reweighting controls indicate that placement matters more than total auxiliary supervision.The effective admission regions for Dpm and Dtm lie at opposite ends of the flow-time axis.Source: §VI · p. 7
  • The lower boundary tpm is data-dependent, whereas the upper boundary ttm is policy-dependent; both are practical offline criteria, not exact task-optimal thresholds.The joint-training results suggest heterogeneous sources can coexist when assigned source-dependent admission regions.Source: §VI · p. 7
  • Experiments are limited to one robot platform, relatively few demonstrations, and a controlled multicamera setup, leaving scalability and cross-embodiment generality incompletely evaluated.The tasks cover both high-precision and coarse-grained manipulation.Source: §VI · p. 7
  • Future work will evaluate ε4P across more robot platforms and hardware configurations and use substantially larger, more diverse robot and human demonstration datasets.These directions address the current limits in platform and demonstration coverage.Source: §VI · p. 7

VII. CONCLUSION

  • Up to 31.7 percentage points over perfect target-task data alone, ε4P improves success rates; replacing equal amounts of additional perfect data incurs an average 4.2-point drop across both imperfect-data sources and all three tasks.The conclusion identifies flow time as a useful axis for integrating heterogeneous imperfect data.Source: §VII · p. 7
Loading 2609.26672v1…