← Paper Digest

Agile Monoped Hopping on Complex and Discrete Terrain via Kinodynamic Planning with Distilled MPPI Steering

Yue Qin, Denglin Cheng, Jingyu Liu, Yanran Ding

November 2026

The simulated monoped hops up a series of inclined platforms into the goal region.

TL;DR FlowRRT plans multi-jump routes for a planar one-legged hopper over gaps, slopes and stepping stones by distilling a multi-start MPPI solver into a flow-matching sampler of all feasible jumps. In simulation it solves 93% of planning problems within 1 s, where SST solves 26% and RL-RRT none.

What they do

A one-legged hopper is severely underactuated, so each foothold must set up the momentum for the next jump. Kinodynamic RRT can plan over many jumps, but every tree extension needs a slow steering step that finds a feasible jump. FlowRRT replaces it with a learned sampler distilled from a trajectory optimizer. Because feasible jumps on discrete terrain form narrow, disconnected sets, the sampler predicts their whole distribution rather than a single answer.

How it works

  • Jump model. The robot is a custom planar monoped (1.97 kg, two 0.22 m leg links) modeled as a single rigid body. Each jump is a 7-D action: flight time, foothold position, stance time and a Bézier ground-reaction-force profile. Friction, torque, joint-limit and collision constraints are checked on top.
  • Expert. Offline, a multi-start MPPI-CEM solver (40 restarts, run on the GPU in NVIDIA Warp) keeps every feasible jump it finds for each steering query.
  • Sampler. A 1.70 M-parameter transformer trained with conditional flow matching, conditioned on the state, the target foothold and the local height map, generates 1024 candidate jumps in 10 Euler steps. Each candidate is checked against the constraints.
  • Search. A learned reachability estimator skips parents that cannot reach the sampled foothold. Nodes that fail repeatedly are pruned, and one steer adds several children by clustering the feasible jumps.

Results

  • On 250 simulated runs per terrain class, FlowRRT solves 99.2% (Gap), 80.8% (Slope) and 100% (Stones) within 1 s. SST reaches less than half of these rates and RL-RRT solves none.
  • Within 1 s it solves at least twice as many runs as the MPPI expert it was distilled from, though the expert reaches a higher final success rate on Slope.
  • Replacing the flow sampler with a single-action regressor drops 1 s success to 42.0% on Slope and 17.2% on Stones. Removing the reachability estimator raises median planning time by 60-67%.
  • A PPO tracking policy trained in Isaac Lab and tested in MuJoCo lands all five planned jumps in all 16 runs, with a mean foothold error of 0.68 cm.

Limitations

Everything is in simulation, including the tracking test (sim-to-sim). Hardware experiments are left to future work. The robot is planar, and on the hardest terrain the distilled sampler misses some feasible jumps the expert finds.

Figures

From Figure 1: Distilled MPPI steering. Left, the multi-start MPPI-CEM expert solves one steering query offline and k...
From Figure 1: Distilled MPPI steering. Left, the multi-start MPPI-CEM expert solves one steering query offline and keeps every feasible action, which form several distinct clusters. Right, a conditional flow matching model maps Gaussian noise to the same clusters in a single online query.
From Figure 3: One FlowRRT search. Parents the reachability estimator predicts cannot reach the sampled foothold are ...
From Figure 3: One FlowRRT search. Parents the reachability estimator predicts cannot reach the sampled foothold are rejected before steering, and nodes that fail repeatedly are deactivated. The extracted path jumps backward onto the lower platform before clearing the gap.
From Figure 4: Success rate versus planning time on the Gap, Slope and Stones terrain classes for FlowRRT, RL-RRT, SS...
From Figure 4: Success rate versus planning time on the Gap, Slope and Stones terrain classes for FlowRRT, RL-RRT, SST and the Expert MPPI reference. Each class has 50 random terrains run with five planner seeds, 250 runs per method; the right edge shows success at the full 100 s budget.
Abstract

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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.