Research map / Large language models / Agents, code and math
Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning
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 NeSyRo, a recursive neuro-symbolic framework that pairs SMT-based code verification with LLM and PDDL-planner-based validation (NeSyConf) and safe exploratory probe code to ground LLM-generated robot code under partial observability.
Problem
LLM code-as-policies for robot task planning become unreliable in dynamic, partially observable settings because generated code is ungrounded and lacks missing observations, leading to incorrect or incomplete code and irreversible actions.
Approach
Combines symbolic verification (Z3 SMT checks of generated code against a task spec) with interactive validation: each skill gets a neuro-symbolic confidence (LLM perplexity-based common-sense confidence times PDDL-planner logic confidence); low-confidence skills trigger recursively verified exploratory probe code that gathers missing observations before the main policy is refined.
Key results
- Across RLBench observability levels, NeSyRo improves success rate by about 26.3% over the strongest baselines (AutoGen, CaP w/ CodeSift) on average, per the paper's text.
- In real-world tasks, NeSyRo reaches 57.5% SR vs 10.6% for CaP (total), with irreversible actions reduced from 53 (CaP) to 7.
- Headline claim: 46.2% task success improvement over Code as Policies baselines and over 86.8% executability of task-relevant actions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add recursive symbolic verification and interactive validation with safe probing | CaP w/ CodeSift · Real-world total, partial observability | SR: 10.6{ \textcolor{gray}{$\pm$4.4}} -> 57.5{ \textcolor{gray}{$\pm$3.5}} | Helped |
| add recursive symbolic verification and interactive validation with safe probing | CaP w/ CodeSift · Real-world total, partial observability | IA: 29 -> 7 | Helped |
| add recursive symbolic verification and interactive validation with safe probing | CaP w/ CodeSift · RLBench, all task types and observability levels | SR: +26.3% | Helped |
| add logic confidence (LC) (ablation) | NeSyRo · RLBench long-horizon, High/Low incompleteness | SR: 44.3{ \textcolor{gray}{$\pm$2.0}} -> 61.9{ \textcolor{gray}{$\pm$6.2}} | Helped |
| add common sense confidence (CSC) (ablation) | NeSyRo · RLBench long-horizon, High/Low incompleteness | SR: 37.1{ \textcolor{gray}{$\pm$4.0}} -> 61.9{ \textcolor{gray}{$\pm$6.2}} | Helped |
| add logic confidence (LC) (ablation) | NeSyRo · RLBench long-horizon, High/Low incompleteness | GC: 54.2{ \textcolor{gray}{$\pm$0.1}} -> 72.0{ \textcolor{gray}{$\pm$3.9}} | Helped |
| add common sense confidence (CSC) (ablation) | NeSyRo · RLBench long-horizon, High/Low incompleteness | GC: 49.9{ \textcolor{gray}{$\pm$4.4}} -> 72.0{ \textcolor{gray}{$\pm$3.9}} | Helped |
| increase CSC model from Llama-3.2-1B to Llama-3.2-3B (ablation) | NeSyRo · RLBench long-horizon, High/Low incompleteness | SR: 32.5{ \textcolor{gray}{$\pm$3.5}} -> 45.0{ \textcolor{gray}{$\pm$0.0}} | Helped |
| use CSC models with 3B or more parameters instead of Llama-3.2-1B (ablation) | NeSyRo · RLBench long-horizon, High/Low incompleteness | SR: 45.0{ \textcolor{gray}{$\pm$0.0}}; 45.0{ \textcolor{gray}{$\pm$7.1}}; 45.0{ \textcolor{gray}{$\pm$0.0}} | No clear effect |
| use GPT-4.1 instead of GPT-4o-mini for code generation | NeSyRo · RLBench long-horizon, High/Low incompleteness | SR: 45.0{ \textcolor{gray}{$\pm$0.0}} -> 50.0{ \textcolor{gray}{$\pm$7.1}} | Helped |
Surprising findings
- Stronger LLMs (o3) do not uniformly benefit the CaP baseline gains; NeSyRo's gains persist across all tested LLMs.
- CaP w/ CodeSift incurs more compile errors than base CaP under partial observability, attributed to hallucinated evaluations without grounded feedback.
Limitations
- Binary logic confidence and predefined domain knowledge limit generality.
- Skills are restricted to those explicitly defined in the domain knowledge.
- Auxiliary manipulation remains low even under complete observation due to light-switch execution accuracy.
Experimental setup
Datasets: RLBench tasks (self-defined task types) real-world Franka tabletop tasks
Benchmarks: RLBench real-world Franka Research 3 tabletop tasks
Metrics: success rate goal condition irreversible actions compile error rate
Baselines: Code as Policies CaP w/ Lemur CaP w/ CodeSift LLM-Planner AutoGen
Builds on: Code as Policies SayCan Chain-of-thought prompting Fast Downward
Closest papers by meaning
- Efficient Skill Grounding via Code Refactoring with Small Language Models (ICML 2026 · similarity 0.91)
- NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks In Open Domains (ICLR 2025 · similarity 0.90)
- CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation (ICML 2026 · similarity 0.90)
- Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making (NeurIPS 2024 · similarity 0.90)
- R2C: Mapping Room to Chessboard to Unlock LLM As Low-Level Action Planner (CVPR 2025 · similarity 0.90)
- NeSyPr: Neurosymbolic Proceduralization For Efficient Embodied Reasoning (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.