Recoleta
GitHub
Source note

PHASER: Phase-Aware and Semantic Experience Replay for Vision-Language-Action Models

Authors
Ziyang Chen, Shaoguang Wang, Weiyu Guo, Qianyi Cai, He Zhang, Pengteng Li, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionContinual LearningExperience ReplayRobot ManipulationCatastrophic Forgetting
Open arXivMarkdown source

Summary

PHASER improves continual learning for vision-language-action robot policies by changing replay memory from uniform frame storage to phase-aware storage and phase-level replay routing. It targets catastrophic forgetting in sequential LIBERO manipulation tasks and reports large ASR gains across 3 VLA backbones.

Problem

  • VLA robot policies forget earlier manipulation tasks when trained on a sequence of new language-conditioned tasks.
  • Uniform experience replay stores frames in proportion to phase length, so short contact-heavy phases such as grasping can get too few examples.
  • Uniform replay also spends memory on old tasks without checking which past phases are most likely to conflict with the current task.

Approach

  • PHASER splits each trajectory into sub-skill phases such as approach, grasp, and transport.
  • Each phase gets an equal frame budget, with stride-3 temporal subsampling and reservoir sampling inside the phase boundary.
  • A replay router scores old phases against the current phase using language similarity, visual similarity, and action divergence.
  • Replay sampling uses a Boltzmann distribution over cached phase scores, computed once at each task transition, so inner-loop replay cost matches standard ER.
  • Auto-PC can replace human phase labels by using action-signal change-point detection followed by VLM semantic verification.

Results

  • On LIBERO-Goal with OpenVLA-OFT-7B, PHASER reaches 87.8% ASR and 7.8 NBT, compared with ER at 77.6% ASR and 22.2 NBT.
  • On LIBERO-Long with OpenVLA-OFT-7B, PHASER reaches 85.8% ASR, compared with ER at 54.6% and MIR at 82.2%.
  • Across the main table, PHASER improves ASR over ER by +10.2 to +39.6 points depending on backbone and suite; the abstract reports up to +31 points over matched-budget ER.
  • On QwenGR00T-3B, PHASER improves LIBERO-Goal ASR from 51.6% to 78.0% and LIBERO-Long ASR from 31.4% to 48.6% versus ER.
  • On QwenOFT-3B, PHASER improves LIBERO-Goal ASR from 39.4% to 79.0% and LIBERO-Long ASR from 33.0% to 51.6% versus ER.
  • Auto-PC phase discovery stays close to human phase labels on LIBERO-Long: 89.6 vs 85.8 ASR on OpenVLA-OFT-7B, 48.0 vs 48.6 on QwenGR00T-3B, and 49.0 vs 51.6 on QwenOFT-3B.

Link