Long-horizon mobile manipulation requires a task plan and the motion that executes it. Generative planners trained on demonstration produce both in one pass, requiring neither a symbolic domain nor search. To date, however, they have relied on thousands of scripted demonstrations of fixed-base arms and executed open loop. This paper presents a hybrid flow matching planner: a single network generates the symbolic plan with masked discrete flow matching and the motion trajectory with continuous flow matching. Unlike prior generative planners, we aim to learn from a much smaller set of demonstrations and to execute the plan in closed loop. Two properties of the data compensate for the small dataset. A demonstration resumed from any of its intermediate actions is itself a demonstration, which multiplies the training samples and enables replanning after every action. Objects of the same kind are interchangeable, which turns demonstrations of one goal into demonstrations of every permuted goal. Our base implementation produces valid plans on 68% of held-out scenes; a training and generation scheme for the discrete plan raises this to 76%, and replanning after every action raises the task completion rate from 40% to 53% in a kinematic simulation. The planner matches the task completion rate of motion-only flow matching policies while additionally providing the symbolic plan, and it outperforms previous hybrid diffusion formulations on both task completion and plan validity. We validate the planner on a real mobile manipulator. Videos and project page: https://andreumatoses.github.io/research/hybrid-flow-planning
Figures & tables
Fig. 1: The planner on the real robot building the tower R on Y on G. Four frames of one run: the start and end of the first half of the task (top) and of the second half (bottom). Each frame shows the plan the planner generated from that scene, with the action in progress highlighted, and the generated waypoints of the base and end effector. The planner generates the waypoints of the whole task; for readability only those of the half shown are drawn.
Fig. 2: The running example. Three cubes R, Y, and G are placed on tables T1–T3, with approach regions RG1–RG10 located along the table sides. The observation o contains the poses of the mobile manipulator and the cubes. A plan consists of a sequence of symbolic actions, such as Pick G , while the goal g specifies the desired final configuration: R on Y on G on T1, illustrated by the ghost cubes.
Fig. 3: The network. The symbolic plan tokens and the waypoints get a token embedding or a linear layer, a learned modality and positional embedding, and form one sequence for the transformer. Its cross-attention reads the memory: one token per flow time (Fourier features) and one per observation component (an MLP), each with a type embedding. The token head gives pθ at every symbolic position and the velocity head vθ at every waypoint.
Fig. 4: Plan suffixes of one demonstration. Top: the end-effector path over the scene, colored per action (blue Move_to , yellow Pick Y and Place Y , red Pick R and Place R ); the grey curves mark the boundary jitter around the action boundaries. Bottom: the plan σ and the suffixes that start at the observations o1,…,oM ; each suffix is a training sample.
Fig. 5: Symbolic permutation of the same demonstration. Renaming the cubes (R to G, Y to R, G to Y) in the goal, the plan, and the observation gives a demonstration of another tower order with the same trajectory.
Fig. 6: The 109 (+12 validation) demonstrations. Left: top view with the tables T1–T3 and the approach regions RG1–RG10: base paths (grey), end-effector paths (violet), start poses (arrows), and initial cube positions (dots). Right: one demonstration represented by observations o1,…,o8 at the action boundaries, together with the corresponding actions, wrist-camera views, and the distribution of plan lengths at the bottom right.
Fig. 7: A run on the real robot with the goal R on Y on G on table T1 (cubes R red, Y yellow, G green). After every action the policy observes the new scene and samples a new plan and trajectory. Here every action succeeds, so the plan stays the same from one planning step to the next and the run looks like an open-loop execution.
Fig. 8: Success against the number of actions left to the goal, on the 12 held-out demonstrations (mean and standard deviation over three training seeds, 50 samples per task). Values of d with few demonstrations are grouped: 13–18, 11–12, and 9–10 hold 850, 750, and 900 tasks, the other values 550 to 600 each. Left: the sampled plan is valid. Middle: the tower stands after the whole sampled trajectory is executed. Right: the tower stands when the policy replans after every action. Hybrid diffusion is the formulation of Høeg et al. [ 13 ] , trained on our suffix samples or on whole demonstrations. Motion-only FM is flow matching on the trajectory alone; it has no plan and no left panel. *The goal-conditioned policy builds any of the six tower orders and is scored on the recorded one.
Symbolic plan valid † (%)
Task completed ‡ (%)
Policy
all d
d≥13
open loop
closed loop
Motion-only FM
–
–
39 ± 2
52 ± 2
Hybrid diffusion [ 13 ] :
whole demonstrations
0 8 ± 1
0 1
0 0
0 0
our suffix samples
62 ± 1
0 5
0 6 ± 0
0 9 ± 0
Ours
76 ± 1
33
40 ± 3
53 ± 3
TABLE I: Performance in percent on the 12 held-out demonstrations, averaged over 3 training seeds and 50 samples per task.
Fig. 9: The same setup with a failure. The grasp of R misses; at the next planning step, the policy sees R still on the table, plans the pick again, and finishes the tower. Replanning after every action turns a missed grasp into a retry instead of a failed run.
Open-loop end-to-end neural motion planners have recently been proposed to improve motion planning for robotic manipulators. These methods enable planning directly from sensor observations without relying on a privileged collision checker during motion planning. However, existing planners produce a single path for a given planning problem and cannot exploit their open-loop nature to propose multiple motion plans. To address this limitation, we introduce Flow Motion Policy, an open-loop neural motion planner that uses flow matching to generate a batch of motion plan proposals by learning a distribution over motion plans conditioned on the planning observation. At inference time, it samples multiple candidate motion plans to enable efficient best-of-N inference while avoiding iterative collision checking during planning. We benchmark the Flow Motion Policy against representative sampling-based, optimization-based and neural motion planning methods. Evaluation results demonstrate that Flow Motion Policy improves planning success and efficiency, highlighting the effectiveness of stochastic generative policies for end-to-end motion planning and best-of-N sampling. Project website: https://davoodsz.github.io/FlowMotionPolicy.github.io/
Davood Soleymanzadeh, Xiao Liang, Minghui Zheng
J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA · Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77843 USA
In recent years Flow Matching has become a prominent method for generative modeling robot motion generation. In its generic form Flow Matching is an ODE-based neural sampler that is trained by regressing empirical flow fields associated with motion samples as data. However, in robot motion generation we often have additional constraints that might not be present in the collected data. The majority of current approaches train the flow on the available data and use inference-time guidance to enforce task-specific constraints. To address this mismatch, we propose ConFlow, a constraint-guided flow matching framework that incorporates constraint information directly into the training objective via differentiable barrier or cost functions. To address design specifications such as smoothness and boundary conditions, we propose replacing the standard Gaussian source distribution used in flow matching training with a conditional Gaussian Process. Our approach also uses infeasible demonstrations as negative supervision, improving constraint satisfaction without requiring additional expert data. Experiments on a two-robot navigation task demonstrate that ConFlow achieves lower collision rates and higher trajectory quality than standard flow matching baselines, with or without inference-time guidance. These results validate training-time constraint integration as an effective approach to closing the training--inference gap in generative motion models.
Nutan Chen, Jianxiang Feng, Marvin Alles +1
LS Wiiri Robot Innovation Center · Independent Researcher · Technical University of Munich +1
Recent advances in generative models have shown promise in generating behavior plans for long-horizon, sparse reward tasks. While these approaches have achieved promising results, they often lack a principled framework for hierarchical decomposition and struggle with the computational demands of real-time execution, due to their iterative denoising process. In this work, we introduce Hierarchical Diffusion-Flow (HDFlow), a novel hierarchical planning framework that optimally leverages the strengths of diffusion and rectified flow models to overcome the limitations of single-paradigm generative planners. HDFlow employs a high-level diffusion planner to generate sequences of strategic subgoals in a learned latent space, capitalizing on diffusion's powerful exploratory capabilities. These subgoals then guide a low-level rectified flow planner that generates smooth and dense trajectories, exploiting the speed and efficiency of ordinary differential equation (ODE)-based trajectory generation. We evaluate HDFlow on four challenging furniture assembly tasks in both simulation and real-world, where it significantly outperforms state-of-the-art methods. Furthermore, we also showcase our method's generalizability on two long-horizon benchmarks comprising diverse locomotion and manipulation tasks. Project website: https://hdflow-page.github.io/
Nandiraju Gireesh, Yuanliang Ju, Chaoyi Xu +3
1Peking University · University of Toronto. · 2Galbot