HUME

Simple Implementation accompanying the Robotics: Science and Systems (RSS) 2026 paper Hypothesis-driven Model Expansion under Uncertainty for Open-World Robot Planning (paper).

This release contains two simulation domains:

• Block Processing World: symbolic manipulation with latent processor effects.
• Mobile Manipulation World (Simulation): household navigation and manipulation in AI2-THOR/ProcTHOR.

This release excludes the real-robot and appliance-operation components because of the substantial systems engineering effort involved. However, most implementation details are provided in the appendix of our paper.

For questions about the real-robot system or potential collaborations, please contact anxingxiao@gmail.com.

HUME requires Python 3.8 or later. Python environments and dependencies are managed with uv:

curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync --all-extras

The lockfile records the complete Python environment for both domains. Run Python commands through uv run; manual virtual-environment activation is unnecessary.

Symbolic variants require Fast Downward. On Ubuntu, install its build prerequisites and run the provided installer:

sudo apt-get install build-essential cmake git
bash scripts/install_fast_downward.sh
export HUME_FAST_DOWNWARD="$PWD/.tools/fast-downward/fast-downward.py"

The installer clones and compiles Fast Downward under .tools/, which is excluded from version control. A different destination may be supplied as its first argument.

Model-backed variants require OPENAI_API_KEY. AI2-THOR and ProcTHOR are installed by uv sync --all-extras; mobile experiments also require a compatible graphics environment.

uv run pytest -q

The focused test suite includes two representative planner cases per domain and solver regression tests. It does not run the complete experiment sweeps.

The five configurations are llm_nh_nv, llm_h_nv, llm_h_v, pddl_h_nv, and pddl_h_v.

A complete experimental sweep for one domain incurs approximately US$100 in API charges.

# Block Processing World
uv run python -m experiments.block_processing.run \
  --planner_type pddl_h_v --model_index normal

# Mobile Manipulation World
uv run python -m experiments.mobile_manipulation.run \
  --planner_type pddl_h_v --model_index normal --start 1

For bounded runs, add --limit N --output-tag TAG; interrupted runs support --resume. Traces, aggregate tables, and figures are written under runs/ and results/. Both directories are excluded from version control.

Generate figures from local experiment outputs with:

uv run python -m experiments.block_processing.plot_results
uv run python -m experiments.mobile_manipulation.plot_results

@inproceedings{xiao2026hypothesis,
  title     = {Hypothesis-driven Model Expansion under Uncertainty for Open-World Robot Planning},
  author    = {Xiao, Anxing and Zhang, Hanbo and Hu, Tianrun and Hsu, David},
  booktitle = {Robotics: Science and Systems (RSS)},
  year      = {2026}
}

This project is released under the MIT License.