Points2Plans: From Point Clouds to Long-Horizon Plans with Composable Relational Dynamics

Code to accompany our ICRA 2025 paper: Points2Plans: From Point Clouds to Long-Horizon Plans with Composable Relational Dynamics. [PDF] [Website]

Approach Overview

This repository includes:

• :hammer_and_wrench: A relational dynamics model that excels at long-horizon prediction of point cloud states without the need to train on multi-step data
• :rocket: A latent-geometric space dynamics rollout strategy that significantly increases the horizons over which predicted point cloud states are reliable for planning
• 🦾 A task planning and goal prediction module using Large Language Models (LLMs)

Setup

System Requirements

This codebase is primarily tested on Ubuntu 20.04, an NVIDIA GeForce RTX 3090 Ti, and CUDA 11.7.

Virtual Env Installation

conda env create -f conda_env.yml

Task Planning and Goal Prediction Module with LLMs

python LLM/scripts/llm_planner.py \ 
    --model-config LLM/configs/models/pretrained/generative/$Model \ 
    --prompt-config LLM/configs/prompts/evaluation/p1/$Task  \
    --api-key $YourAPIKey

Relational Dynamics

Quick Start with Pretrained Models

• Download pretrained models from this link
• Download test data for constrained packing task

python relational_dynamics/main.py \
    --result_dir $PretrainedModelDir  \
    --checkpoint_path $PretrainedModelDir/checkpoint/pretrained.pth \ 
    --test_dir $TestDataDir  \
    --test_max_size $TestSize 

Training

• Download training datasets

To generate your own data, please refer to our simulation repository using [isaacgym].

python relational_dynamics/main.py \
    --result_dir $YourResultDir  \
    --train_dir $TrainingDataDir \
    --batch_size $BatchSize \
    --num_epochs $TrainingEpochs \
    --max_size $TrainingSize 

Baseline: eRDTransformer

Training

python relational_dynamics/main.py \
    --result_dir $YourResultDir  \
    --train_dir $TrainingDataDir \
    --batch_size $BatchSize \
    --num_epochs $TrainingEpochs \
    --delta_forward False \ 
    --latent_forward True \ 
    --max_size $TrainingSize 

Test

• Download pretrained models from this link

python relational_dynamics/main.py \
    --result_dir $PretrainedModelDir  \
    --checkpoint_path $PretrainedModelDir/checkpoint/baseline_pretrained.pth \ 
    --test_dir $TestDataDir  \
    --delta_forward False \ 
    --latent_forward True \ 
    --test_max_size $TestSize 

Citation

If you find our work useful in your research, please cite:

@InProceedings{huang-icra2025-p2p,
 title = {{Points2Plans: From Point Clouds to Long-Horizon Plans with Composable Relational Dynamics}},
 author = {Yixuan Huang and Christopher Agia and Jimmy Wu and Tucker Hermans and Jeannette Bohg},
 booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
 year = {2025},
 url = {https://sites.google.com/stanford.edu/points2plans}
}