Trust the PRoC3S: Solving Long-Horizon Robotics Problems with LLMs and Constraint Satisfaction

MIT CSAIL

Place 5 objects in a line

Place all objects into same-color bowls

Pack all objects within 0.1m of the table center

Abstract

Recent developments in pretrained large language models (LLMs) applied to robotics have demonstrated their capacity for sequencing a set of discrete skills to achieve open-ended goals in simple robotic tasks. In this paper, we examine the topic of LLM planning for a set of continuously parameterized skills whose execution must avoid violations of a set of kinematic, geometric, and physical constraints. We prompt the LLM to output code for a function with open parameters, which, together with environmental constraints, can be viewed as a Continuous Constraint Satisfaction Problem (CCSP). This CCSP can be solved through sampling or optimization to find a skill sequence and continuous parameter settings that achieve the goal while avoiding constraint violations. Additionally, we consider cases where the LLM proposes unsatisfiable CCSPs, such as those that are kinematically infeasible, dynamically unstable, or lead to collisions, and re-prompt the LLM to form a new CCSP accordingly. Experiments across three different simulated 3D domains demonstrate that our proposed strategy, PRoC3S, is capable of solving a wide range of complex manipulation tasks with realistic constraints on continuous parameters much more efficiently and effectively than existing baselines.



Real World Constraints

Grasp Failures

Grasp Failures

Obstacle Collisions

Obstacle Collisions

Kinematic Infeasibility

Kinematic Infeasibility

Lack of Stability

Lack of Stability



The PRoC3S

An LLM is prompted with an example initial state, goal, LMP, and associated LMP domain for drawing a square. When prompted with a new state and goal for drawing a star, the language model outputs a new LMP and associated domain. We then sample inputs to the function and test them against a set of pre-specified constraints via a simulator. If no satisfying assignment is found after N samples, we feed back the primary failure modes to the LLM to generate an updated LMP and domain.

PRoC3S


Drawing Domain

Drawing Star

Draw a star

Drawing Star

Draw a rectangle enclosing any obstacle

Drawing Star

Draw the letter M

Drawing Star

Draw an arrow pointing at the largest obstacle





Blocks Domain

Drawing Star

Place 5 objects in a line

Drawing Star

Pack all objects into the region

Drawing Star

Stack a pyramid of blocks

Drawing Star

Place a green block in a green bowl





YCB Domain

Drawing Star

Stack any two objects

Drawing Star

Pack all objects witin 0.1m of the table center



BibTeX


@misc{curtis2024trustproc3ssolvinglonghorizon,
      title={Trust the PRoC3S: Solving Long-Horizon Robotics Problems with LLMs and Constraint Satisfaction}, 
      author={Aidan Curtis and Nishanth Kumar and Jing Cao and Tomás Lozano-Pérez and Leslie Pack Kaelbling},
      year={2024},
      eprint={2406.05572},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2406.05572}, 
}