RobotDiffuse: Diffusion-Based Motion Planning for Redundant Manipulators with the ROP Obstacle Avoidance Dataset

TL;DR

The paper tackles motion planning for redundant manipulators in complex, obstacle-rich environments. It introduces RobotDiffuse, a diffusion-model-based planner that integrates physical constraints via a CAE-based obstacle encoder and uses an encoder-only Transformer to capture long-range temporal dependencies, improving planning coherence. A large-scale PyBullet-based ROP dataset with 35M poses and 0.14M obstacle-avoidance scenarios is released to benchmark diffusion-based planners. Empirical results show RobotDiffuse achieves high success rates quickly, surpasses baselines on the proposed ComScore, and demonstrates diffusion models' promise for robust, efficient motion planning, with the dataset made publicly available for future work.

Abstract

Redundant manipulators, with their higher Degrees of Freedom (DoFs), offer enhanced kinematic performance and versatility, making them suitable for applications like manufacturing, surgical robotics, and human-robot collaboration. However, motion planning for these manipulators is challenging due to increased DoFs and complex, dynamic environments. While traditional motion planning algorithms struggle with high-dimensional spaces, deep learning-based methods often face instability and inefficiency in complex tasks. This paper introduces RobotDiffuse, a diffusion model-based approach for motion planning in redundant manipulators. By integrating physical constraints with a point cloud encoder and replacing the U-Net structure with an encoder-only transformer, RobotDiffuse improves the model's ability to capture temporal dependencies and generate smoother, more coherent motion plans. We validate the approach using a complex simulator and release a new dataset, Robot-obtalcles-panda (ROP), with 35M robot poses and 0.14M obstacle avoidance scenarios. The highest overall score obtained in the experiment demonstrates the effectiveness of RobotDiffuse and the promise of diffusion models for motion planning tasks. The dataset can be accessed at https://github.com/ACRoboT-buaa/RobotDiffuse.

Figures (3)

• Figure 1: Schematic of the motivation in this paper. Comparing the sample-based and our diffusion-based approaches. $S$ and $T$ denote the number of sampling and diffusion steps, respectively. Sample-based methods require more time to generate a path and often involve frequent backtracking.
• Figure 2: RobotDiffuse overview. It contains three steps: embedding, training, and generation. (1) The obstacle point net is embedded into the latent space as $z$ by the PointEnet component. Together with the init pose $CS_{init}$, the target pose $CS_{goal}$ comprises a condition code $C$. (2) $C$ is randomly masked for classifier-free learning and then projected with $t$ into the input token $z_{tk}$. Given a noise and motion $X$, the motion predictor based on an encoder-only transformer predicts the original clean motion $\hat{X}^{1:N}$. (3) Given a condition $C$, we sample random noise $X_T$ at the dimensions of the desired motion, then iterate from $T$ to $1$. At each step $t$, the predictor predicts a clean sample $\hat{X}$ and then adds noise, bringing it back to $X_{t-1}$.
• Figure 3: Comparison of success rates over time and visualization of robot obstacle avoidance trajectories.