Generalizable Motion Planning via Operator Learning
Authors: Sharath Matada, Luke Bhan, Yuanyuan Shi, Nikolay Atanasov
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work, we introduce a planning neural operator (PNO) for predicting the value function of a motion planning problem. We recast value function approximation as learning a single operator from the cost function space to the value function space, which is defined by an Eikonal partial differential equation (PDE). Therefore, our PNO model, despite being trained with a finite number of samples at coarse resolution, inherits the zero-shot super-resolution property of neural operators. We demonstrate accurate value function approximation at 16 the training resolution on the Moving AI lab s 2D city dataset, compare with state-of-the-art neural value function predictors on 3D scenes from the i Gibson building dataset and showcase optimal planning with 4-DOF robotic manipulators. Lastly, we investigate employing the value function output of PNO as a heuristic function to accelerate motion planning. We show theoretically that the PNO heuristic is ϵ-consistent by introducing an inductive bias layer that guarantees our value functions satisfy the triangle inequality. With our heuristic, we achieve a 30% decrease in nodes visited while obtaining near optimal path lengths on the Moving AI lab 2D city dataset, compared to classical planning methods (A , RRT ). |
| Researcher Affiliation | Academia | Sharath Matada , Luke Bhan , Yuanyuan Shi, Nikolay Atanasov University of California, San Diego EMAIL |
| Pseudocode | No | The paper describes the PNO architecture and its components (lifting, Fourier layers, projection) in detail with explanations and Figure 2, but does not include a separate, formally labeled pseudocode or algorithm block. |
| Open Source Code | Yes | See project code https://github.com/Existential Robotics/PNO. |
| Open Datasets | Yes | Moving AI lab s 2D city dataset (Sturtevant, 2012), compare with state-of-the-art neural value function predictors on 3D scenes from the i Gibson building dataset and showcase optimal planning with 4-DOF robotic manipulators. |
| Dataset Splits | Yes | In the first experiment, we consider a small Grid-World dataset as in Tamar et al. (2016), consisting of 5k training maps and 1k testing maps. We compare with VIN and IEF2D. For planning, we perform gradient descent on the test-map value function predictions from VIN, IEF2D, and PNO. |
| Hardware Specification | Yes | To clarify, all the 2D experiments use a NVIDIA 3090 Ti. The 3D i Gibson experiments use a NVIDIA A100 for data-generation and training, while we employ a NVIDIA 4060 during testing. |
| Software Dependencies | No | The paper mentions several software tools like MATLAB, Pykonal FMM, and SciPy for numerical solvers, but it does not provide specific version numbers for any of these components or other libraries that would be necessary to replicate the experiments. |
| Experiment Setup | Yes | To develop our model, it took approximately 40 minutes of training on a NVIDIA RTX 3090 Ti GPU. [...] Our architecture consisted of 157808 parameters and we achieved an L2 relative error of 0.1 for both training and testing taking approximately 10 minutes for training on a NVIDIA 3090 Ti GPU. [...] For comparison, we train NTFields and P-NTFields models on the Bolton and Samuel environments which takes approximately 3 hours to train per environment (NVIDIA A100 GPU). [...] The model was trained and tested on a NVIDIA 3090Ti GPU. [...] For training, FMM was used and the data generation took 40s while the model took 20 min to be trained over 200 epochs. [...] For this example, the model was trained with ξ = 0.05 in (15). |