NeuPAN

Real-time end-to-end planner that maps raw lidar points directly to robot actions

License

GPL-3.0

Overview

NeuPAN (Neural Proximal Alternating-minimization Network) is an end-to-end, real-time, map-free robot navigation planner published in IEEE Transactions on Robotics. It couples a learned point-to-collision distance model with model-based optimization solved by proximal alternating minimization, mapping raw lidar points directly to collision-free actions — no perception-to-planning pipeline, no prebuilt map — and it runs on real differential-drive and Ackermann robots in cluttered, dynamic environments.

Conventional navigation stacks chain perception, mapping, and planning modules, accumulating latency and error at each stage. NeuPAN collapses the chain: an interpretable optimization problem couples robot dynamics, control limits and obstacle-avoidance constraints, where obstacle distances come from DUNE, a learned model that encodes raw points into latent distance features. The problem is solved online by a neural proximal alternating-minimization network, giving real-time rates on embedded GPUs while retaining the constraint-satisfaction behaviour of model-based control. The repository provides the full planner with pretrained models, YAML-configured examples in IR-SIM (the author's lightweight simulator, also on this platform), ROS integration for real robots, and reproductions of the paper's scenarios including dynamic obstacles, narrow corridors and dense unstructured clutter with differential-drive and Ackermann kinematics.

Why it matters

NeuPAN is one of the first end-to-end planners with both a rigorous model-based backbone and demonstrated real-robot deployments — published in T-RO, it represents where navigation research is heading: direct sensor-to-action planning with the safety structure of optimization, not just a black-box policy.

Problem it solves

Modular navigation pipelines add latency and compounding error between perception and planning, and struggle in dense, dynamic clutter. NeuPAN plans directly on raw points in real time while enforcing dynamics and collision constraints, closing that gap on real hardware.

Key takeaways

• End-to-end: raw lidar points map directly to collision-free actions — no prebuilt map
• Model-based core: dynamics, control limits and collision constraints in one optimization
• DUNE learned distance model + proximal alternating-minimization solver for real-time rates
• Deployed on real differential-drive and Ackermann robots in dynamic clutter
• Published in IEEE Transactions on Robotics; simulation examples ship on IR-SIM

Who it is for

• Navigation researchers exploring end-to-end planning
• Robotics engineers evaluating learned planners
• Graduate students studying optimization-based control
• Developers of robots in cluttered dynamic spaces

Limitations

• GPL-3.0 license restricts closed-source productization
• Requires PyTorch and benefits from a GPU for real-time performance
• 2D navigation focus — no 3D flight or manipulation planning
• Research code: integration beyond the provided ROS examples is on the user