Dynamic Neural Potential Field: Online Trajectory Optimization in Presence of Moving Obstacles
Aleksei Staroverov, Muhammad Alhaddad, Aditya Narendra, Konstantin Mironov, Aleksandr Panov
AI summary
Problem
Mobile robots struggle to plan safe, smooth trajectories in real-time when obstacles move unpredictably, as classical MPC requires fast, differentiable collision models that are difficult to derive from raw dynamic occupancy maps.
Approach
The authors propose NPField-GPT, a learning-enhanced MPC framework that uses a non-autoregressive Transformer to predict a horizon of footprint-aware repulsive potentials from occupancy grids, which are then injected into a sequential quadratic MPC solver via L4CasADi.
Key results
- Real-time MPC-compatible dynamic neural potential framework with three accuracy-latency variants
- Novel NPField-GPT architecture predicting horizon potentials in parallel with enhanced obstacle-relative geometry conditioning
- Superior safety and trajectory efficiency over CIAO* and MPPI baselines in simulations and real Husky UGV experiments
- Open-source implementation with L4CasADi integration for differentiable neural costs inside MPC
Why it matters
Enables generalist robots to navigate safely and efficiently in unpredictable human environments by combining the robustness of model-based planning with data-driven spatial-temporal risk estimation.
Abstract
Generalist robot policies must operate safely and reliably in everyday human environments such as homes, offices, and warehouses, where people and objects move unpredictably. We present Dynamic Neural Potential Field (NPField-GPT), a learning-enhanced model predictive control (MPC) framework that couples classical optimization with a Transformer-based predictor of footprint-aware repulsive potentials. Given an occupancy sub-map, robot footprint, and optional dynamic- obstacle cues, our NPField-GPT model forecasts a horizon of differentiable potentials that are injected into a sequen- tial quadratic MPC program via L4CasADi, yielding real- time, constraint-aware trajectory optimization. We additionally study two baselines: NPField-StaticMLP, where a dynamic scene is treated as a sequence of static maps; and NPField- DynamicMLP, which predicts the future potential sequence in parallel with an MLP. In dynamic indoor scenarios from BenchMR and on a Husky UGV in office corridors, NPField-GPT produces more efficient and safer trajectories under motion changes, while StaticMLP/DynamicMLP offer lower latency. We also com- pare with the CIAO* and MPPI baselines. Across meth- ods, the Transformer+MPC synergy preserves the trans- parency and stability of model-based planning while learn- ing only the part that benefits from data: spatiotempo- ral collision risk. Code and trained models are avail- able at https://github.com/CognitiveAISystems/ Dynamic-Neural-Potential-Field.