Venue: ICLR 2026 Authors: Shiying Duan, Pei Ren, Nanxiang Jiang, Zhengping Che, Jian Tang, Zhaoxin Fan, Yifan Sun, Wenjun Wu arXiv: 2506.06683 Category: Policy learning — task/motion planning for dual-arm robots Trend tag: LLM task planning / bimanual parallelism

flowchart LR
  Task[Multi-task instruction<br/>dual-arm scenario] --> LLM[LLM-driven planner]
  LLM --> S1[Stage 1: Dependency-Graph-based<br/>Planning Candidates Generation<br/>build DAGs, remove redundancy]
  S1 --> DAG[Directed Acyclic Graphs<br/>of task dependencies]
  DAG --> S2[Stage 2: Graph Re-Traversal-based<br/>Dual-Arm Parallel Planning<br/>optimize traversal for max parallelism]
  S2 --> Plan[Parallel dual-arm plan<br/>left + right arm schedules]
  Plan --> Exec[Execution]
  XDAPT[X-DAPT dataset<br/>cross-scenario eval] -.evaluates.-> Plan

Dual-arm robots can improve efficiency and flexibility in complex multitasking, but existing task-planning methods often fail to fully optimize task parallelism, leaving the two arms underused. The challenge is to model inter-task dependencies and then schedule subtasks across both arms so as to maximize concurrency without violating coherence/ordering constraints.

RoboPARA is an LLM-driven framework with a two-stage pipeline:

1. Dependency Graph-based Planning Candidates Generation — constructs directed acyclic graphs (DAGs) that model task dependencies and eliminate redundancy among subtasks.
2. Graph Re-Traversal-based Dual-Arm Parallel Planning — optimizes DAG traversal to maximize parallelism between the two arms while maintaining task coherence.

The authors also introduce X-DAPT (Cross-Scenario Dual-Arm Parallel Task dataset), described as the first dataset specifically designed to evaluate dual-arm task parallelism across diverse scenarios and difficulty levels.

Across the X-DAPT scenarios, RoboPARA reportedly outperforms existing planning methods on efficiency and reliability, with the largest gains on complex task combinations. (The paper frames improvements qualitatively; specific aggregate numbers are not restated here.)

RoboPARA targets the planning layer of bimanual manipulation rather than the low-level policy: it treats parallelism as a graph-optimization problem over an LLM-generated dependency structure, and contributes a dedicated benchmark (X-DAPT) for measuring dual-arm concurrency. This complements policy-level bimanual work that focuses on coordinated low-level control.

• arXiv: https://arxiv.org/abs/2506.06683
• OpenReview: https://openreview.net/forum?id=b9GH5mMBfG
• Code: https://github.com/AiDuanshiving/RoboPARA

• ICLR 2026 Survey
• VLA Architectures review
• Dexterous Manipulation review

← Back to ICLR-2026