Research map / Reinforcement learning / MDPs and planning

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

NeurIPS 2026 · Reinforcement learning › MDPs and planning · approach: POMDP policy optimization · 0 citations (OpenAlex)

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.

Introduces PO-PDDL, a symbolic POMDP formulation, and a pipeline that learns PO-PDDL domains from real-robot visual demonstrations for online DESPOT belief-space planning.

Problem

Robot task planning under both stochastic action execution and partial observability requires POMDP models, which are labor-intensive to build; existing PDDL learners assume deterministic full observability and existing POMDP learners are not evaluated on visual robot domains.

Approach

A symbolic POMDP language (PO-PDDL) extending PDDL with stochastic effects, observable schemas, action-conditioned symbolic observation rules, and belief specifications; the domain is learned in two stages by reconstructing hidden symbolic state trajectories from videos and detecting partial observability from contradictions between reconstructed states and visual parses.

Key results

• PO-PDDL achieves 100% task success across all five real-world tasks with 10 trials each.
• PO-PDDL domain learning takes 291 s versus 5111 s for UniDomain* and 328 s for pomdp_coder*.
• PO-PDDL planning time is on average 50.4% lower than VLM-Planner, 85.7% lower than UniDomain*, and 73.2% lower than pomdp_coder*.

Surprising findings

• UniDomain* and pomdp_coder* fail on the long-horizon tasks, and pomdp_coder* succeeds only on the single-step push_box task.

Limitations

• Assumes a closed domain without post-demonstration domain shift.
• Learning cost may grow with large-scale demonstration datasets.
• Performance bounded by the capability and biases of the underlying VLM.

Experimental setup

Datasets: 29 real-world visual demonstrations (103 steps)

Benchmarks: push_box block_in_drawer cup_with_water cups_then_blocks box_cups_blocks

Metrics: task progress cumulative reward planning time success rate

Baselines: VLM-Planner UniDomain* pomdp_coder*

Builds on: PDDL PPDDL POMDPX DESPOT UniDomain pomdp_coder

Closest papers by meaning

• UniDomain: Pretraining a Unified PDDL Domain from Real-World Demonstrations for Generalizable Robot Task Planning (NeurIPS 2025 · similarity 0.91)
• Tru-POMDP: Task Planning Under Uncertainty via Tree of Hypotheses and Open-Ended POMDPs (NeurIPS 2025 · similarity 0.89)
• Simplifying Complex Observation Models in Continuous POMDP Planning with Probabilistic Guarantees and Practice (AAAI 2024 · similarity 0.89)
• RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation (NeurIPS 2025 · similarity 0.87)
• Hierarchical Planning and Learning for Robots in Stochastic Settings Using Zero-Shot Option Invention (AAAI 2024 · similarity 0.87)
• Learning POMDP World Models from Observations with Language-Model Priors (NeurIPS 2026 · similarity 0.86)