Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond the Majority: Long-tail Imitation Learning for Robotic Manipulation

Junhong Zhu, Ji Zhang, Jingkuan Song, Lianli Gao, Heng Tao Shen

TL;DR

This work tackles the long-tail data distribution problem in imitation learning for generalist robotic manipulation policies, showing that tail tasks suffer from degraded spatial reasoning due to data scarcity. It identifies phase-specific failure, particularly in the target-approaching phase, as the main bottleneck and proposes Approaching-Phase Augmentation (APA), which transfers head-task approaching-phase knowledge to tail tasks by head-to-tail object grafting and linguistically aligned co-training. The authors validate APA with extensive simulation on LIBERO-Core-LT and real-world experiments (Real-World-LT), demonstrating significant gains for tail tasks without sacrificing head-task performance. The results highlight APA’s practical appeal, its robustness to task-distribution changes, and its potential to make broad, data-efficient generalist policies more reliable in real-world manipulation scenarios.

Abstract

While generalist robot policies hold significant promise for learning diverse manipulation skills through imitation, their performance is often hindered by the long-tail distribution of training demonstrations. Policies learned on such data, which is heavily skewed towards a few data-rich head tasks, frequently exhibit poor generalization when confronted with the vast number of data-scarce tail tasks. In this work, we conduct a comprehensive analysis of the pervasive long-tail challenge inherent in policy learning. Our analysis begins by demonstrating the inefficacy of conventional long-tail learning strategies (e.g., re-sampling) for improving the policy's performance on tail tasks. We then uncover the underlying mechanism for this failure, revealing that data scarcity on tail tasks directly impairs the policy's spatial reasoning capability. To overcome this, we introduce Approaching-Phase Augmentation (APA), a simple yet effective scheme that transfers knowledge from data-rich head tasks to data-scarce tail tasks without requiring external demonstrations. Extensive experiments in both simulation and real-world manipulation tasks demonstrate the effectiveness of APA. Our code and demos are publicly available at: https://mldxy.github.io/Project-VLA-long-tail/.

Beyond the Majority: Long-tail Imitation Learning for Robotic Manipulation

TL;DR

This work tackles the long-tail data distribution problem in imitation learning for generalist robotic manipulation policies, showing that tail tasks suffer from degraded spatial reasoning due to data scarcity. It identifies phase-specific failure, particularly in the target-approaching phase, as the main bottleneck and proposes Approaching-Phase Augmentation (APA), which transfers head-task approaching-phase knowledge to tail tasks by head-to-tail object grafting and linguistically aligned co-training. The authors validate APA with extensive simulation on LIBERO-Core-LT and real-world experiments (Real-World-LT), demonstrating significant gains for tail tasks without sacrificing head-task performance. The results highlight APA’s practical appeal, its robustness to task-distribution changes, and its potential to make broad, data-efficient generalist policies more reliable in real-world manipulation scenarios.

Abstract

While generalist robot policies hold significant promise for learning diverse manipulation skills through imitation, their performance is often hindered by the long-tail distribution of training demonstrations. Policies learned on such data, which is heavily skewed towards a few data-rich head tasks, frequently exhibit poor generalization when confronted with the vast number of data-scarce tail tasks. In this work, we conduct a comprehensive analysis of the pervasive long-tail challenge inherent in policy learning. Our analysis begins by demonstrating the inefficacy of conventional long-tail learning strategies (e.g., re-sampling) for improving the policy's performance on tail tasks. We then uncover the underlying mechanism for this failure, revealing that data scarcity on tail tasks directly impairs the policy's spatial reasoning capability. To overcome this, we introduce Approaching-Phase Augmentation (APA), a simple yet effective scheme that transfers knowledge from data-rich head tasks to data-scarce tail tasks without requiring external demonstrations. Extensive experiments in both simulation and real-world manipulation tasks demonstrate the effectiveness of APA. Our code and demos are publicly available at: https://mldxy.github.io/Project-VLA-long-tail/.
Paper Structure (23 sections, 6 equations, 6 figures, 7 tables)

This paper contains 23 sections, 6 equations, 6 figures, 7 tables.

Table of Contents

  1. INTRODUCTION
  2. Related works
  3. Generalist Robot Policies
  4. Learning from Long-tail Distributions
  5. Diagnosing the Long-tail Failure Mode
  6. Problem formulation
  7. A Long-tail Imitation Learning Benchmark
  8. Inefficiency of Re-sampling Methods
  9. Phase-Wise Failure Analysis for Long-Tail Learning
  10. Approaching-Phase Augmentation (APA)
  11. EXPERIMENTS
  12. Simulation Experiments
  13. Experimental Setup
  14. Results and Analysis
  15. Ablation Studies
  16. ...and 8 more sections

Figures (6)

  • Figure 1: Influence of long-tail demonstrations on policy generalization. Owing to the skewed distribution of training demonstrations, the performance degradation of generalist robot policies is particularly pronounced for tail tasks relative to those in the head.
  • Figure 2: Comparison of training demonstration distribution between the LIBERO-Core-FULL (blue bar) and LIBERO-Core-LT (yellow bar) datasets.
  • Figure 3: Overview of the Approaching-Phase Augmentation (APA) pipeline. Our method involves a three-step process: (1) Head Task Trajectory Segmentation, which isolates the target approaching phase from data-rich head tasks; (2) Tail to Head Object Grafting, which creates augmented trajectories using objects from tail tasks; and (3) Instruction Formatting and Co-Training, which formats the corresponding language instructions and then trains the policy on the combined dataset.
  • Figure 4: Effectiveness of our proposed APA method on LIBERO-Core-LT dataset.
  • Figure 5: Effectiveness of our proposed APA method on real-word long-tail tasks.
  • ...and 1 more figures