BridgeFlow: Fast and Robust SE(2)-Equivariant Motion Planning with Flow Matching

Robotic motion planning requires spatial generalization—a trajectory planned for one configuration should transfer to translated or rotated versions of the same task. Current learning-based planners face an unfortunate trade-off: they either lack built-in equivariance (treating every spatial transformation as a new problem) or enforce it through computationally expensive specialist architectures that limit real-time inference. BridgeFlow breaks this trade-off with a principled combination of canonicalization and flow matching.

Core Innovation

Rather than building an equivariant neural network, BridgeFlow achieves exact SE(2) equivariance through a lightweight task-centric canonicalization module that transforms the input task into a canonical coordinate frame before planning. This allows the use of standard, unconstrained neural architectures for trajectory generation while inheriting strict equivariance guarantees.

For inference acceleration, the framework pairs a Brownian bridge informative prior with context-aware mini-batch optimal transport, constructing a straightened vector field that minimizes transport cost and stabilizes training. Classifier-Free Guidance explicitly embeds environmental awareness into the generation process, ensuring the planned trajectories respect obstacles and constraints.

Results

Evaluation in dense 2D environments and on a 7-DoF Franka Emika Panda manipulator demonstrates:

• Up to 15× inference speedup over state-of-the-art diffusion-based motion planners
• 2× higher valid trajectory rate, meaning far fewer infeasible or colliding plans
• Robust generalization to entirely unseen environment layouts
• Exact equivariance under arbitrary SE(2) transformations of the task configuration
• Training stability improvements from the optimal-transport-based vector field straightening

Limitations

The framework is currently limited to SE(2) equivariance; extension to full SE(3) for 6-DoF manipulation tasks is left as future work. The lightweight canonicalization approach relies on the existence of well-defined canonical frames for the task geometry, which may not always be straightforward for unstructured environments. Evaluation on the Franka manipulator was conducted in simulation.

Industry Implications

Motion planning is a daily bottleneck for industrial robot arms in manufacturing, logistics, and assembly. The combination of strict equivariance guarantees with standard (non-specialized) architectures means BridgeFlow’s speed and generalization benefits are accessible without rewriting existing neural planners. The 15× speedup brings high-quality generative planning closer to real-time control rates, which is particularly valuable for applications requiring rapid re-planning in response to changing environments—such as bin picking in cluttered warehouses or collaborative assembly cells.