Research map / Multi-agent systems and games / LLM agents

RoboPARA: Dual-Arm Robot Planning with Parallel Allocation and Recomposition Across Tasks

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RoboPARA, an LLM-driven two-stage framework (DAG generation with validation/correction, then graph re-traversal scheduling) for dual-arm task parallelism planning, plus the X-DAPT benchmark.

Problem

Dual-arm robots executing multiple concurrent task packages are typically planned sequentially by LLM agents, leaving arm parallelism and collaboration underexploited.

Approach

Models multi-task instructions as a dependency DAG generated by an LLM with RAG-retrieved procedural knowledge and iterative rule-based correction, then re-traverses the DAG with a deadlock-aware scheduler that assigns single- and dual-arm operations to maximize parallelism.

Key results

• On X-DAPT with GPT-4o, RoboPARA reaches TEI 1.956 on easy packages and 0.869 on hard packages with TFR 0.041, versus the best baseline TFR of 0.111 on hard packages.
• RoboPARA's parallel step ratio (PPR) is over 10x that of baselines on medium and hard packages according to the paper's claims.
• The paper reports 30% to 50% reduction in execution time and 34% higher success rate than the average of other methods on challenging tasks.

Surprising findings

• VOYAGER fails to produce concrete execution steps in all scenes (TFR 1.000), and RoCo is almost unable to solve any task after more than 50 prompting iterations.
• Removing instructional constraints in the ablation raises failure rate substantially on medium packages (TFR 0.742 vs 0.000), while removing DAG correction increases it moderately.

Limitations

• High API cost: RoboPARA uses on average 1.3x more tokens than baselines due to iterative DAG correction.
• Accuracy on complex packages with intricate dependencies remains imperfect within limited correction rounds.
• Skill library is tightly coupled to predefined scenarios and package templates, limiting generalization to unseen tasks.

Experimental setup

Datasets: X-DAPT RoboTwin

Benchmarks: X-DAPT

Metrics: TEI TFR PPR APR

Baselines: LLM^3 ChatGPT-Prompts VOYAGER Embodied TaPA LLM-Planner FLTRNN RoCo

Builds on: RoCo FLTRNN LLM-Planner retrieval augmented generation

Closest papers by meaning

• RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins (CVPR 2025 · similarity 0.88)
• UniDomain: Pretraining a Unified PDDL Domain from Real-World Demonstrations for Generalizable Robot Task Planning (NeurIPS 2025 · similarity 0.86)
• RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation (NeurIPS 2025 · similarity 0.85)
• Embodied Task Planning via Graph-Informed Action Generation with Large Language Models (ICML 2026 · similarity 0.85)
• MRBTP: Efficient Multi-Robot Behavior Tree Planning and Collaboration (AAAI 2025 · similarity 0.85)
• ManiTaskGen: A Comprehensive Task Generator for Benchmarking and Improving Vision-Language Agents on Embodied Decision-Making (NeurIPS 2026 · similarity 0.85)