Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Achieving Scalable Robot Autonomy via neurosymbolic planning using lightweight local LLM

Nicholas Attolino, Alessio Capitanelli, Fulvio Mastrogiovanni

TL;DR

Gideon introduces a neurosymbolic planning framework that runs on lightweight local LLMs to enable scalable, on-device PDDL planning for robot autonomy. A key novelty is an automated Domain-Problem Generation Configuration and data pipeline that generates large-scale, multi-domain domain-problem-plan tuples compatible with PDDL 2.1, enabling on-premise training and evaluation. Empirical results using Qwen-2.5 1.5B show up to 66.1% valid plans for a 32k single-domain model and 70.6% in multi-domain settings, with cross-domain data variety improving generalization. The work demonstrates the feasibility of cross-domain, on-device neurosymbolic planning with extended context windows, offering predictable latency and adaptability for human-robot collaboration, while acknowledging slower training/inference and the need for further optimization and domain expansion.

Abstract

PDDL-based symbolic task planning remains pivotal for robot autonomy yet struggles with dynamic human-robot collaboration due to scalability, re-planning demands, and delayed plan availability. Although a few neurosymbolic frameworks have previously leveraged LLMs such as GPT-3 to address these challenges, reliance on closed-source, remote models with limited context introduced critical constraints: third-party dependency, inconsistent response times, restricted plan length and complexity, and multi-domain scalability issues. We present Gideon, a novel framework that enables the transition to modern, smaller, local LLMs with extended context length. Gideon integrates a novel problem generator to systematically generate large-scale datasets of realistic domain-problem-plan tuples for any domain, and adapts neurosymbolic planning for local LLMs, enabling on-device execution and extended context for multi-domain support. Preliminary experiments in single-domain scenarios performed on Qwen-2.5 1.5B and trained on 8k-32k samples, demonstrate a valid plan percentage of 66.1% (32k model) and show that the figure can be further scaled through additional data. Multi-domain tests on 16k samples yield an even higher 70.6% planning validity rate, proving extensibility across domains and signaling that data variety can have a positive effect on learning efficiency. Although long-horizon planning and reduced model size make Gideon training much less efficient than baseline models based on larger LLMs, the results are still significant considering that the trained model is about 120x smaller than baseline and that significant advantages can be achieved in inference efficiency, scalability, and multi-domain adaptability, all critical factors in human-robot collaboration. Training inefficiency can be mitigated by Gideon's streamlined data generation pipeline.

Achieving Scalable Robot Autonomy via neurosymbolic planning using lightweight local LLM

TL;DR

Gideon introduces a neurosymbolic planning framework that runs on lightweight local LLMs to enable scalable, on-device PDDL planning for robot autonomy. A key novelty is an automated Domain-Problem Generation Configuration and data pipeline that generates large-scale, multi-domain domain-problem-plan tuples compatible with PDDL 2.1, enabling on-premise training and evaluation. Empirical results using Qwen-2.5 1.5B show up to 66.1% valid plans for a 32k single-domain model and 70.6% in multi-domain settings, with cross-domain data variety improving generalization. The work demonstrates the feasibility of cross-domain, on-device neurosymbolic planning with extended context windows, offering predictable latency and adaptability for human-robot collaboration, while acknowledging slower training/inference and the need for further optimization and domain expansion.

Abstract

PDDL-based symbolic task planning remains pivotal for robot autonomy yet struggles with dynamic human-robot collaboration due to scalability, re-planning demands, and delayed plan availability. Although a few neurosymbolic frameworks have previously leveraged LLMs such as GPT-3 to address these challenges, reliance on closed-source, remote models with limited context introduced critical constraints: third-party dependency, inconsistent response times, restricted plan length and complexity, and multi-domain scalability issues. We present Gideon, a novel framework that enables the transition to modern, smaller, local LLMs with extended context length. Gideon integrates a novel problem generator to systematically generate large-scale datasets of realistic domain-problem-plan tuples for any domain, and adapts neurosymbolic planning for local LLMs, enabling on-device execution and extended context for multi-domain support. Preliminary experiments in single-domain scenarios performed on Qwen-2.5 1.5B and trained on 8k-32k samples, demonstrate a valid plan percentage of 66.1% (32k model) and show that the figure can be further scaled through additional data. Multi-domain tests on 16k samples yield an even higher 70.6% planning validity rate, proving extensibility across domains and signaling that data variety can have a positive effect on learning efficiency. Although long-horizon planning and reduced model size make Gideon training much less efficient than baseline models based on larger LLMs, the results are still significant considering that the trained model is about 120x smaller than baseline and that significant advantages can be achieved in inference efficiency, scalability, and multi-domain adaptability, all critical factors in human-robot collaboration. Training inefficiency can be mitigated by Gideon's streamlined data generation pipeline.
Paper Structure (13 sections, 3 figures, 4 tables)

This paper contains 13 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Planning domains, PDDL version and planner
  5. Training pipeline
  6. Domain-Problem Generation Configuration
  7. Training and testing
  8. Results and discussion
  9. Single-domain
  10. Multi-domain
  11. Conclusion
  12. Acknowledgments.
  13. Disclosure of Interests.

Figures (3)

  • Figure 1: A diagram of Gideon's training pipeline. In order, selected domains are processed in accordance to the rules specified in the DPCG file to generate a set amount of problem instance; then plans are generated by the Plan Generator module with the selected planner; finally, domain-problem-plans tuples are assembled and formatted into the training, validation and test set files.
  • Figure 2: Planning validity achieved by Gideon-NO-MACRO and Gideon-MD with respect to increasing training steps. Gideon-NO-MACRO-8k, -16k, and -32k are reported in orange. For Gideon MD, both results on the multi-domain test set (in green) and on a second single-domain test set (NO-MACRO, in blue) are reported to facilitate comparison with single-domain models. The red line above represent the success rate of the baseline traditional planner Probe (i.e., 100%)
  • Figure 3: Box plots of the planning times of Gideon-NO-MACRO-16k, Gideon-MD-16k tested on 500 NO-MACRO samples, Gideon-MD-16k tested on the complete multi-domain test set, and the baseline traditional planner Probe.