Safe Streaming Flow Planning by Aligning Sampling Dynamics with Execution Dynamics
SafeStreamingFlow is a preprint describing a start–goal-conditioned flow matching planner that streams one-step state updates in physical execution time, so safety filtering only needs to act on the step actually executed. In simulations, the safe version reaches 88% success with 4% violations on Maze2D Large (vs 13% success and 71% violations for SafeDiffuser) and runs Hopper safety filtering with 7.30 ms total latency vs 1012.4 ms for SafeDiffuser. The result matters because it brings execution-time safety filters to generative planners without the usual O(H*K) compute and distribution-shift failures.
Can a generative flow-matching planner enforce execution-time safety constraints and run in real time by aligning trajectory sampling with the physical execution dynamics, rather than refining an entire horizon and filtering unexecuted intermediate states?
Generative diffusion/flow planners synthesize long-horizon trajectories but typically refine all H states over K denoising/sampling iterations; safety corrections applied across these H*K intermediate states are computationally heavy, can produce dynamically inconsistent or locally trapped plans, and because sampled intermediate states are not physically executed, the learned sampling dynamics differ from the system's execution dynamics.
Earlier methods include trajectory-level diffusion planners such as Diffuser and SafeDiffuser, flow matching planners such as FlowMatcher/SafeFlowMatcher, classifier-guidance variants, and classical RRT*/A* with explicit cost maps. Trajectory-level safe planners filter whole horizons with control barrier functions at every sampling step, causing O(H*K) QP solves, distribution shift, local traps, and high latency; streaming flow methods existed mainly as action policies and did not directly handle start–goal planning or execution dynamics.
SafeStreamingFlow trains a start–goal-conditioned state vector field with contracting conditional flow matching around expert demonstrations. It integrates this field forward in physical time one step at a time using Euler updates and a hierarchical state predictor that predicts the next position without seeing current velocity, then predicts the next velocity conditioned on that position, suppressing a kinematic shortcut. The one-step-ahead reference is tracked by a PD controller; the nominal PD action is minimally adjusted by a discrete-time high-order control barrier function QP/QCQP only for the state that will actually be executed, and the resulting safe state is fed back for the next streaming update.
Evaluated over 100 rollouts per environment. Maze2D: SafeStreamingFlow closed-loop reached 98/95/88% success on Umaze/Medium/Large with execution violations 0/0/4% and zero traps; SafeDiffuser reached 67/49/13% with violations 33/46/71% and traps 24/0/18%. Unsafe StreamingFlow closed-loop reached 100/94/91% but violated 45/41/50%. Classical RRT*/A* reached 91/94/86 and 93/85/85 with 5/1/8 and 7/14/6% violations. Optimization solves reduced from O(HK) to O(H), about 5.4x faster solve time on Maze2D. F1TENTH: SafeStreamingFlow reached 73% success on Budapest with 20% violation and 99% on Catalunya with 0% violation, while SafeDiffuser/SafeFM/SafeFlowMatcher all had 0% success on Budapest; unsafe StreamingFlow had 77% success but 97% violation. Hopper: SafeStreamingFlow score 51.4±3.4 with 0% execution violations and 7.30 ms total latency; StreamingFlow 57.4±5.5; SafeDiffuser 16.6±13.7 with 85% violations and 1012.4 ms total. Warehouse navigation: SafeStreamingFlow closed-loop 51% success, 1% violation, 0.07 s optimization time, 2.58 s total; SafeDiffuser 34% success, 19% violation, 6.91 s optimization; unsafe FM 58% success but 18% violation; RRT*/A* 56/41% success with 0 violations.
The authors state that SafeStreamingFlow transports data to data rather than noise to data, so it fundamentally requires richer demonstration data; it inherits CBF limitations such as quadratic barriers restricting obstacle geometry and requiring pre-specified valid barriers and gains; recursive feasibility is not established, and on Maze2D Large the filter occasionally fails with a few closed-loop rollouts violating the barrier; all evaluations are simulation-only. Evident limitations include no real-robot validation, a preprint without peer review, and use of a selected PD controller and hand-specified constraints across simulated environments.
The closest users are autonomous mobile robot and warehouse automation developers using ROS 2/Nav2, because the method runs inside a Nav2/PD control loop without a cost map and reduces safe-planner latency. It could also support safety-critical generative planning for racing or test platforms and legged or humanoid locomotion where control barrier filters are used. Production use is likely 1–3 years away; prerequisites include real-robot validation, recursive feasibility or fallback strategies, support for non-quadratic barrier geometries, richer demonstration data, and integration with perception and state estimation.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.