Curious Now

Story

Graph-of-Constraints Model Predictive Control for Reactive Multi-agent Task and Motion Planning

Artificial IntelligenceComputing

Key takeaway

Researchers developed a new algorithm that helps robots coordinate tasks and movement more efficiently. This could enable better real-world robotic applications like self-driving cars or factory robots working together.

Read the paper

Quick Explainer

The key idea of this work is a novel Graph-of-Constraints (GoC) model that generalizes prior optimization-based approaches to multi-agent task and motion planning. The GoC represents partially-ordered tasks and enables dynamic agent assignments, overcoming limitations of fixed task sequencing and static agent-task mappings. The accompanying GoC-MPC algorithm decomposes the optimization problem into efficient subproblems that can be solved in real-time, enabling reactive execution and handling of disturbances. This reactive, optimization-based approach outperforms a state-of-the-art baseline in simulation and on a physical robotic system, demonstrating strong scalability as the number of agents and objects increases.

Deep Dive

Technical Deep Dive: Graph-of-Constraints Model Predictive Control for Reactive Multi-agent Task and Motion Planning

Overview

This work presents a novel approach called Graph-of-Constraints Model Predictive Control (GoC-MPC) for solving reactive multi-agent Task and Motion Planning (TAMP) problems. GoC-MPC builds upon prior work on optimizing sequences of geometric constraints, but extends this to handle partially-ordered tasks and dynamic agent assignment.

The key contributions are:

  • Graphs-of-Constraints (GoC): A generalization of sequences-of-constraints that can naturally represent partially-ordered tasks and dynamic agent assignments.
  • GoC-MPC: An efficient algorithm that decomposes the GoC optimization problem into subproblems that can be solved in real-time, enabling reactive execution.
  • Experimental validation showing GoC-MPC outperforms a state-of-the-art baseline (ReKep) in terms of success rate, computation time, and path length across simulated and physical multi-agent manipulation tasks.

Problem & Context

  • Effective deployment of multi-robot teams can automate complex real-world tasks, but introduces significant challenges in TAMP.
  • Prior optimization-based TAMP methods using sequences-of-constraints struggle with:
    • Imposing a total order on task steps, preventing parallelism.
    • Static assignment of agents to tasks, unable to adapt to disturbances.

Methodology

Graphs-of-Constraints (GoC)

  • Generalize sequences-of-constraints to a Directed Acyclic Graph (DAG) structure.
  • Allows representing partially-ordered tasks and dynamic agent assignments.
  • Constraints are defined over a dynamic agent assignment matrix.

GoC-MPC Algorithm

  1. Waypoints and Assignments Subproblem:
    • Optimize waypoints and agent assignments jointly to resolve constraints dependent on assignments.
  2. Agent Splines Subproblem:
    • Compute smooth agent trajectories passing through the optimized waypoints.
  3. Short Receding-Horizon Subproblem:
    • Compute a short-horizon trajectory tracking the agent splines while considering fine dynamics.

Reactive Execution

  • GoC-MPC iteratively solves the 3 subproblems, enabling forward phase progression and backtracking through the GoC.
  • Handles disturbances by updating the set of active constraints and re-optimizing.

Data & Experimental Setup

  • Evaluated GoC-MPC on 3 simulated multi-agent manipulation tasks:
    • Block Stacking
    • Pick-and-Pour
    • Tablecloth Folding
  • Compared to state-of-the-art baseline ReKep in both static and disturbance settings.
  • Also evaluated scalability to varying numbers of agents and objects.
  • Validated real-world performance on a dual-UR5e robotic system.

Results

Static Settings

  • GoC-MPC outperformed ReKep in success rate, computation time, and path length across block stacking and pick-and-pour tasks.
  • Key advantages of GoC formulation:
    • Allows parallelism between agents, avoiding unnecessary delays.
    • Optimizes agent assignments jointly, preventing failures from poor static assignments.

Disturbance Settings

  • GoC-MPC achieved 100% success rate on block stacking with disturbances, matching ReKep.
  • However, GoC-MPC's agent-independent backtracking reduced unnecessary movements by 0.64 meters on average and was 80x faster in maximum time.

Scalability

  • As the number of objects and agents increased, GoC-MPC maintained high success rates, low computation times, and short path lengths.

Real-world Validation

  • GoC-MPC achieved comparable performance on physical dual-UR5e system as in simulation, succeeding in nearly all trials.
  • Only failures were due to severe block occlusions during block stacking.

Interpretation

  • The generalized GoC formulation effectively addresses the limitations of prior sequences-of-constraints approaches.
  • GoC-MPC's decomposition and reactive execution strategy enables real-time performance, even under disturbances.
  • The approach demonstrates strong scalability and transfers well to physical robotic systems.

Limitations & Uncertainties

  • Performance can degrade if the external state estimation module is unreliable.
  • The initial GoC plan skeleton assumed in the experiments may not always be available in practice.

What Comes Next

  • Explore integrating learned perception modules to overcome state estimation limitations.
  • Investigate ways to combine GoC-MPC with discrete planning to handle cases where the initial plan skeleton is unavailable.

Source