Reproduce “SeqVLA: Sequential Task Execution for Long-Horizon Manipulation with Completion-Aware Vision-Language-Action Model” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ran Yang, Zijian An, Lifeng ZHou, Yiming Feng

A completion-aware extension to Vision-Language-Action (VLA) models, SeqVLA, integrates a learned subtask completion detection head to enable autonomous transitions in long-horizon robotic manipulation tasks. The system demonstrated a 100% success rate on a 7-subtask salad packing task and 60% on a 4-subtask candy packing task, outperforming a baseline VLA model by eliminating sequence-related failures.

Key insights

  • Adapting the pre-trained VLM backbone during finetuning improves performance, suggesting the importance of domain-specific adaptation for deeper architectural components.
  • Jointly finetuning the action and completion detection heads leads to more confident and statistically distinct subtask completion predictions (lower entropy, higher KS statistic).
  • A tightly coupled, learned subtask completion signal allows VLA models to autonomously manage sequential task execution, mitigating common sequence-related failures.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.