This paper presents FlowRRT, a kinodynamic planner for agile monoped hopping on discrete terrain. Steering for kinodynamic RRT on complex dynamical systems currently relies on motion primitives or closed-loop policies, limiting the diversity of motions or the planning speed. Distilling a trajectory optimizer avoids both, but existing methods regress a single solution per query, which fails on monoped hopping, where the feasible action set is narrow and disconnected. Our key insight is to treat steering as set-valued and predict the distribution of feasible actions instead. A multi-start MPPI expert collects the entire feasible set of each query offline, and conditional flow matching distills it into a generative sampler. With this sampler, FlowRRT coordinates successive jumps over long horizons and finds non-greedy maneuvers that are hard for reactive policies to discover. Benchmarks across different terrains show higher success rates and substantially shorter planning times than sampling-based baselines. On terrains with inclined platforms, FlowRRT reaches 80% success within 1~s, versus 24~s for RL-RRT and 77~s for SST. Sim-to-sim tracking in MuJoCo confirms accurate execution under full-order dynamics.
Figures & tables
Fig. 1: Overview of the distilled MPPI steering. Left, the multi-start MPPI-CEM expert solves one steering query offline and collects every feasible action it encounters, which form several distinct clusters. Right, a conditional flow matching model trained on these actions maps Gaussian noise to the same clusters in a single online query, preserving the multimodality of the expert data.
Fig. 2: Architecture of the conditional velocity field vϕτ(uτ;c) . The conditions are embedded into one state–target token and one token per cell of the local height map, each carrying its position relative to the target foothold and the takeoff foot, and encoded by self-attention blocks. The noised action uτ is embedded as a single token and refined by decoder blocks, where it provides the queries and the encoded context the keys and values. The flow time τ modulates the decoder through adaLN-Zero, and a linear head outputs the velocity.
Fig. 3: One FlowRRT search, following Algorithm 2 . A target foothold is sampled, and parents predicted by the reachability estimator to be unable to reach it are rejected before steering. The selected parent is steered toward the target, which either yields new tree edges or fails, and a node that fails repeatedly is deactivated. Once a node reaches the goal region, the path is extracted, here jumping backward onto the lower platform before clearing the gap.
Fig. 4: Success rate versus planning time for FlowRRT, RL-RRT [ 9 ] , and SST [ 14 ] baselines, and the Expert MPPI reference across terrain classes. For each terrain class, all planners are evaluated on the same 50 randomly generated terrain instances spanning x∈[−1,6] m, with the robot starting at x=0 and the goal region defined as Pgoal=[5,6] m. Each terrain is evaluated with five planner seeds, yielding 250 runs per method per terrain class. At time t , the success rate is the fraction of these runs that have found a valid solution within t seconds, and the value at the right edge of each panel is the success rate at the full budget of 100 s.
Fig. 5: Planned motions on two terrains. Robots mark the contacts of the returned path, drawn at the end of each stance. The solid curve is the CoM path, blue in flight and orange in stance, the dashed curve is the foot path, and arrowheads mark the direction of takeoff velocity. (a) Slope terrain. Arriving too fast to climb after the third jump, the robot steps back onto the flat platform and returns to the slope with a slower approach and a higher foothold for the climb. (b) Stones terrain. The path crosses stepping stones no wider than 0.24 m with one contact per stone.
Fig. 6: Ablation of FlowRRT on the same runs as Fig. 4 . W/o reach removes the reachability estimator, and regressor steering replaces the flow sampler with a regressor, with everything else kept the same. Left, success rate within 1 s in solid color and within 10 s in light color. Right, failed steering ratio, the share of steering calls over all runs that yield no feasible child node.
Fig. 7: Tracking experiment using a learned tracking policy. (a) Simulation snapshots, with solid and dashed lines indicating the simulated and FlowRRT planned CoM trajectories, respectively. (b) State and action trajectories during execution.
Reaching a goal state in a complex environment often requires a robot to overcome challenging terrain and obstacles. Which routes are accessible depends on the robot's physical design, the control policy it uses, and how it modifies its surroundings. We present a kinodynamic rapidly-exploring random tree (RRT) planner that uses physics simulation to exploit skills encoded in the control policy, which allows it to find a path in the scene with respect to the robot's capabilities. The planner samples states from a reduced state space in 3D around the scene and expands the RRT toward each sample using policy rollouts. We guide sampling with a preplanned path for a simple mass-point model to focus exploration on promising regions leading toward the goal state, reducing planning time by up to 63%. Simulation experiments with quadrupeds, humanoids, and wheeled robots demonstrate that our approach supports drop-in integration of various robots and control policies, exploiting their physical and policy capabilities without modifying the underlying planning algorithm. Predefined environmental modifications, such as placing a plank across a gap to build a bridge, will automatically be picked up by the planner as the robot's reachable space expands.
Yael Fassbind, Valentin N. Hartmann, Huang Su +2
Computational Robotics Lab, ETH Zurich, Switzerland.
Robots that leave structured, static environments must plan motions that are kinodynamically feasible and safe among moving obstacles. However, there are no dedicated benchmark frameworks that combine both aspects. To overcome this, we present DynoFluxBench, a framework to compare kinodynamic planners in known, dynamic environments with unbounded arrival time. To demonstrate its utility and establish strong baselines, we develop three dedicated planners, named ST-Db-RRT, ST-GBRRT, and KIST, that fuse kinodynamic and space-time methods, covering different kinodynamic search paradigms: ST-Db-RRT expands with randomly selected discontinuity-bounded motion primitives using trajectory optimization, whereas KIST and ST-GBRRT maintain a kinodynamically feasible tree with different heuristic guidance. We analyze the probabilistic completeness guarantees of those new planners in dynamic environments. Finally, we evaluate ST-Db-RRT, ST-GBRRT, and KIST using DynoFluxBench, showing that ST-Db-RRT reaches a first solution up to 32 times faster, while KIST and ST-GBRRT remain valuable where trajectory optimization is fragile. Videos and further analysis can be found at https://dynofluxbench.github.io/dynofluxbench/.
Quadruped robots can traverse low obstacles, but many 2D planning pipelines still model obstacles as binary occupied regions and rely on sampling-based search that can be inefficient under a limited budget. We propose a perception-assisted height-adaptive planning framework based on CMP-IRRT*, a Channel Mamba PointNet-guided Informed RRT* planner. Given a calibrated top-view RGB observation, the perception module estimates obstacle regions and converts depth predictions into a ground-relative height map. The planner then performs height-conditioned collision checking, treating high obstacles as blocked while allowing low obstacles to be traversed, and uses the CMP guide to bias sampling toward promising regions while retaining standard free-space and informed sampling fallbacks. Experiments on 2D planning benchmarks show that CMP-IRRT* reduces explored nodes and iterations compared with classical and neural-guided baselines, and a controlled ablation supports the contribution of the Mamba-based guide. In constructed traversability-aware scenarios, the proposed planner reduces path length by up to 16.3% when low obstacles are traversable, and a Unitree Go2 demonstration further shows executable bypassing and traversal behaviors. Our code is publicly available at https://github.com/MingfanZhao/height-adaptive-planner.
Mingfan Zhao, Wendong Mao, Zhongfeng Wang
School of Integrated Circuits, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China