Neural MP

Neural MP is a machine learning-based motion planning system for robotic manipulation tasks. It combines neural networks trained on large-scale simulated data with lightweight optimization techniques to generate efficient, collision-free trajectories. Neural MP is designed to generalize across diverse environments and obstacle configurations, making it suitable for both simulated and real-world robotic applications. This repository contains the model weights for Neural MP.

All Neural MP checkpoints, as well as our codebase are released under an MIT License.

For full details, please read our paper and see our project page.

Model Summary

Developed by: The Neural MP team consisting of researchers from Carnegie Mellon University.
Language(s) (NLP): en
License: MIT
Pretraining Dataset: Coming soon
Repository: https://github.com/mihdalal/neuralmotionplanner
Paper: Coming soon
Project Page & Videos: https://mihdalal.github.io/neuralmotionplanner/

Installation

Please read here for detailed instructions

Usage

Neural MP model takes in 3D point cloud and start & goal angles of the Franka robot as input, and predict 7-DoF delta joint actions. We provide a wrapper class NeuralMP for inference and deploy our model in the real world.

Here's an deployment example with the Manimo Franka control library:

Note: using Manimo is not required, you may use other Franka control libraries by creating a wrapper class which inherits from FrankaRealEnv (see franka_real_env.py)

import argparse
import numpy as np
from neural_mp.envs.franka_real_env import FrankaRealEnvManimo
from neural_mp.real_utils.neural_motion_planner import NeuralMP
if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--mdl_url",
        type=str,
        default="mihdalal/NeuralMP",
        help="hugging face url to load the neural_mp model",
    )
    parser.add_argument(
        "--cache-name",
        type=str,
        default="scene1_single_blcok",
        help="Specify the scene cache file with pcd and rgb data",
    )
    parser.add_argument(
        "--use-cache",
        action="store_true",
        help=("If set, will use pre-stored point clouds"),
    )
    parser.add_argument(
        "--debug-combined-pcd",
        action="store_true",
        help=("If set, will show visualization of the combined pcd"),
    )
    parser.add_argument(
        "--denoise-pcd",
        action="store_true",
        help=("If set, will apply denoising to the pcds"),
    )
    parser.add_argument(
        "--train-mode", action="store_true", help=("If set, will eval with policy in training mode")
    )
    parser.add_argument(
        "--tto", action="store_true", help=("If set, will apply test time optimization")
    )
    parser.add_argument(
        "--in-hand", action="store_true", help=("If set, will enable in hand mode for eval")
    )
    parser.add_argument(
        "--in-hand-params",
        nargs="+",
        type=float,
        default=[0.1, 0.1, 0.1, 0.0, 0.0, 0.1, 0.0, 0.0, 0.0, 1.0],
        help="Specify the bounding box of the in hand object. 10 params in total [size(xyz), pos(xyz), ori(xyzw)] 3+3+4.",
    )
    args = parser.parse_args()
    env = FrankaRealEnvManimo()
    neural_mp = NeuralMP(
        env=env,
        model_url=args.mdl_url,
        train_mode=args.train_mode,
        in_hand=args.in_hand,
        in_hand_params=args.in_hand_params,
        visualize=True,
    )
    points, colors = neural_mp.get_scene_pcd(
        use_cache=args.use_cache,
        cache_name=args.cache_name,
        debug_combined_pcd=args.debug_combined_pcd,
        denoise=args.denoise_pcd,
    )
    # specify start and goal configurations
    start_config = np.array([-0.538, 0.628, -0.061, -1.750, 0.126, 2.418, 1.610])
    goal_config = np.array([1.067, 0.847, -0.591, -1.627, 0.623, 2.295, 2.580])
    if args.tto:
        trajectory = neural_mp.motion_plan_with_tto(
            start_config=start_config,
            goal_config=goal_config,
            points=points,
            colors=colors,
        )
    else:
        trajectory = neural_mp.motion_plan(
            start_config=start_config,
            goal_config=goal_config,
            points=points,
            colors=colors,
        )
    success, joint_error = neural_mp.execute_motion_plan(trajectory, speed=0.2)