PO-PDDL: Learning Symbolic POMDPs from Visual Demonstrations for Robot Planning Under Uncertainty

2026-06-14 · Source: Artificial Intelligence · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

PO-PDDL introduces a symbolic formulation for Partially Observable Markov Decision Processes (POMDPs) designed for real-world robot task planning under stochastic action execution and partial observability. This new language preserves the relational structure and LLM-friendly syntax of PDDL, while explicitly modeling beliefs, stochasticity, and partial observability. The method includes a demonstration-driven pipeline that reconstructs latent symbolic state trajectories from real-robot execution videos. It identifies partial observability through inconsistencies between inferred states and visual observations, then learns stochastic transition and observation models. Experiments on long-horizon manipulation tasks show PO-PDDL consistently outperforms existing PDDL and POMDP model-learning approaches, enabling robust task planning with significantly lower planning cost.

Key takeaway

For robotics engineers and AI scientists building autonomous systems that operate under uncertainty, PO-PDDL offers a robust approach to constructing POMDP models. This method allows you to learn complex planning domains directly from visual demonstrations, significantly reducing the labor involved in model creation. Consider adopting PO-PDDL to enhance your robot's ability to perform long-horizon tasks reliably, even with perception and execution uncertainties, while also lowering planning costs.

Key insights

PO-PDDL learns symbolic POMDPs from visual demonstrations for robust robot planning under uncertainty.

Principles

• Symbolic POMDPs can preserve PDDL's relational structure.
• Partial observability is identifiable via state-observation inconsistencies.
• Demonstration-driven learning improves POMDP model construction.

Method

Reconstruct latent symbolic state trajectories from robot videos, identify partial observability via state-observation inconsistencies, then learn stochastic transition and observation models.

In practice

• Use PO-PDDL for robot planning under uncertainty.
• Apply visual demonstrations to learn POMDP models.
• Leverage PO-PDDL for reusable task domains.

Topics

• Robotics
• POMDPs
• PDDL
• Symbolic AI
• Machine Learning
• Visual Demonstrations
• Uncertainty Planning

Best for: Research Scientist, Robotics Engineer, AI Scientist

Related on AIssential

• Scaling Observation-aware Planning in Uncertain Domains — Decomposing POMDPs enables vastly more scalable solutions for optimal sensor selection in uncertain domains.
• Google Goes to EXTREMES: In-Context Symbolic AI — LLMs can perform symbolic reasoning with high accuracy when rigorously constrained by in-context PDDL definitions.
• Generative adversarial imitation learning for robot swarms: Learning from human demonstrations and trained policies — A GAIL framework enables robot swarms to learn collective behaviors from human demonstrations, performing comparably to source policies.
• Learning Visual Spatial Planning from Symbolic State via Modality-Gap-Aware Self-Distillation — MGSD bridges the perception-reasoning modality gap in visual planning through a two-stage self-distillation framework.
• Learning-guided Prioritized Planning for Lifelong Multi-Agent Path Finding in Warehouse Automation — Integrating RL with search-based planning significantly enhances lifelong multi-agent pathfinding in complex environments.
• From Kinematics to Dynamics: Learning to Refine Hybrid Plans for Physically Feasible Execution — Robotic plans based on first-order dynamics are physically infeasible; a neuro-symbolic RL framework can refine them for second-order feasibility.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.