FlashPlan: A Fast Neural Motion Planner for Robotic Ball Catching

Highlights

• A single-pass neural motion planner that generates collision-free, time-normalized joint-space trajectories with low and consistent latency — a median of ~6.5 ms versus 90–160 ms for classical sampling-based planners.
• A torque-constrained time-scaling formulation that computes the fastest dynamically feasible execution time for the planned motion via monotonicity-based binary search.
• Planning latency that stays low and spatially uniform across the workspace, unlike classical planners whose latency depends on the specific intercept geometry.
• Raises simulated ball-catching success from 87.4% (RRT-Connect) and 82.8% (EST) to 95.2%, while more than doubling the available control window (330.6 ms vs. ~230 ms).

Method

Overview

In contrast to prior learning-based planners that guide or warm-start iterative search, FlashPlan eliminates search entirely and formulates motion generation as a feedforward prediction problem. Given the current joint configuration, the goal configuration, and an environment representation, FlashPlan directly predicts a compact, collision-free, time-normalized joint-space trajectory in a single forward pass — low and consistent latency, independent of obstacle complexity. A torque-constrained time-scaling stage then computes the fastest execution time for that trajectory that keeps every joint within its torque limits, and a torque-based low-level controller tracks the resulting time-scaled trajectory.

Fig. 2 — The planning and control pipeline. An IK solver produces q_goal, which together with q_init and the obstacle representation feeds FlashPlan’s context and map networks. The output network predicts feasibility, timing, and joint path-points, which are torque-constrained time-scaled and tracked by a low-level controller.

Trajectory Representation

FlashPlan outputs a compact, time-normalized joint-space trajectory represented by a small, fixed number of path-points {(s_i, q_i)}, where s_i ∈ [0, 1] is a normalized time and q_i ∈ R⁷ a joint configuration; a continuous trajectory is reconstructed by interpolating between them. Decoupling motion generation from absolute execution time this way lets the same predicted geometry be scaled to whatever duration is dynamically feasible. Training labels come from RRTConnect expert demonstrations, compressed to a handful of path-points with the Ramer–Douglas–Peucker algorithm.

Fig. 3 — An example time-normalized joint-space trajectory represented with a small number of path-points (circular markers); the continuous trajectory is recovered by interpolation.

Similar to prior learning-based planners such as MPNet, FlashPlan also includes a binary feasibility head that predicts whether a collision-free path exists between q_init and q_goal given the environment — approximating the classical feasibility query in constant time.

Network Architecture

FlashPlan separately encodes a structured context representation (initial and goal joint configurations plus obstacle center/bounding-box descriptions, as a fixed-size 2D tensor) and a 2D top-down occupancy grid of the workspace. A context network (a convolutional stem followed by residual blocks) and a map network (convolutional blocks interleaved with downsampling, followed by adaptive average pooling) each produce a fixed-dimensional embedding; the two are concatenated and passed to a shared output network — an MLP backbone with three heads predicting (i) a feasibility score, (ii) normalized path-point times {s_i}, and (iii) the corresponding joint configurations {q_i}.

FlashPlan is trained entirely offline on 10,000 samples generated by treating RRTConnect as an expert in randomized cluttered environments (80/10/10 train/val/test split), with a weighted multi-task loss over feasibility, path-point timing, joint-space accuracy, goal-reaching, and velocity smoothness.

Torque-Constrained Time Scaling

Because the predicted trajectory is time-normalized, an execution duration T must be chosen before it can be tracked. Joint velocities scale as 1/T and accelerations as 1/T², so the peak fraction of each joint’s torque limit used along the trajectory, Φ(T), decreases monotonically as T increases. FlashPlan exploits this monotonicity to binary-search for the fastest torque-feasible execution time, T^* = min{T : Φ(T) ≤ 1}, with negligible runtime overhead — maximizing the control window left before the ball arrives while guaranteeing the motion stays within the manipulator’s dynamic limits.

Limitations

FlashPlan is trained offline on trajectories from a classical planner (RRTConnect) and therefore inherits that expert’s structural biases. Transferring the method to a different robot requires retraining due to differences in physical dimensions. And while torque-constrained time scaling guarantees dynamic feasibility of the planned motion, real tracking performance still depends on the low-level controller and hardware limitations. The present study also assumes the interception pose is estimated upstream; closing the loop with direct high-speed visual perception is left to future work.

Citation

The paper is currently under review at ICRA 2027. If you’d like to reference it in the meantime:

@misc{chowdhury2026flashplan,
  title  = {FlashPlan: A Fast Neural Motion Planner for Robotic Ball
            Catching},
  author = {Chowdhury, Sakib and Guo, Yi},
  note   = {Submitted to ICRA 2027},
  year   = {2026}
}