Reproduce “Distilling On-device Language Models for Robot Planning with Minimal Human Intervention” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zachary Ravichandran, Ignacio Hounie, Fernando Cladera, Alejandro Ribeiro, George J. Pappas, Vijay Kumar
The PRISM framework distills powerful cloud-based Large Language Models (LLMs) into efficient Small Language Models (SLMs) that operate directly on robotic hardware for planning tasks. This approach enables on-device SLMs to achieve planning success rates comparable to frontier LLMs while significantly reducing latency and eliminating cloud dependency for robot control.
Key insights
- High-performance, cloud-based LLM planning capabilities can be effectively distilled into compact, on-device SLMs that maintain comparable task success rates.
- Automated data synthesis, through generative LLMs and a closed-loop planning emulation with environment masking and plan validation, is crucial for creating robust training datasets without manual curation.
- On-device deployment of distilled SLMs provides significantly faster and more deterministic planning execution (under 200ms query times) compared to cloud-dependent LLMs, crucial for real-time robotic applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.