Decoupled diffusion policies
Dedicated arm and base policies learn in their own observation and action spaces.
Robot learning
Loco-manipulation
FALCON: Actively Decoupled Visuomotor Policies for Loco-Manipulation with Foundation-Model-Based Coordination
National University of Singapore * Equal contribution
From visual understanding to coordinated whole-body action.
Research overview & real-world demonstrations
01 / The idea
Independent policies.
Shared understanding.
Moving and manipulating require different perspectives. FALCON gives each its own policy, then brings them together with a vision-language foundation model.
Specialized diffusion policies act on their own observations. A shared representation of the scene, instruction, and task progress helps the arm and base work together.
Read the full abstract
We present FoundAtion-model-guided decoupled LoCO-maNipulation visuomotor policies (FALCON), a framework for loco-manipulation that combines modular diffusion policies with a vision–language foundation model as the coordinator. Our approach explicitly decouples locomotion and manipulation into two specialized visuomotor policies, allowing each subsystem to rely on its own observations. This mitigates the performance degradation that arises when a single policy is forced to fuse heterogeneous, potentially mismatched observations from locomotion and manipulation. Our key innovation lies in restoring coordination between these two independent policies through a vision–language foundation model, which encodes global observations and language instructions into a shared latent embedding conditioning both diffusion policies. On top of this backbone, we introduce a phase-progress head that uses textual descriptions of task stages to infer discrete phase and continuous progress estimates without manual phase labels. To further structure the latent space, we incorporate a coordination-aware contrastive loss that explicitly encodes cross-subsystem compatibility between arm and base actions. We evaluate FALCON on two challenging loco-manipulation tasks requiring navigation, precise end-effector placement, and tight base-arm coordination. Results show that it surpasses centralized and decentralized baselines while exhibiting improved robustness and generalization to out-of-distribution scenarios.
02 / The method
Specialized control.
Semantic coordination.
A modular architecture connects local observations with a global understanding of the task.
Task-aware coordination
A foundation model provides shared context, with phase and progress inferred without manual phase labels.
Compatible whole-body actions
A coordination-aware contrastive objective aligns the two subsystems in the shared latent space.
03 / In the real world
Coordination in action.
Explore the task demonstrations, onboard observations, and additional robustness experiments.
Place the toy in the drawer, then close it.
Coordinated navigation, object placement, and drawer closing.
Human-in-the-loop control
Control one subsystem while the learned policy operates the other. These demonstrations explore the modularity of FALCON’s decoupled policies.
Human-controlled locomotion
Human-controlled manipulation
Open the drawer, pick the toy, and place it inside.
A multi-stage sequence linking precise manipulation with base motion.
Human-in-the-loop control
Control one subsystem while the learned policy operates the other. These demonstrations explore the modularity of FALCON’s decoupled policies.
Human-controlled locomotion
Human-controlled manipulation
Generalization to different starting positions
The robot starts from multiple locations across the workspace. Training demonstrations are primarily collected in the center region; left and right regions test out-of-distribution initial positions.
Hang the light brown hat on the yellow hook.
Precise placement with varying object positions and external disturbances.
Additional placement trial
Response to external disturbances
Pick up the red parcel and place it in the yellow box.
Whole-body manipulation under varied placements and external disturbances.
Additional placement trial
Response to external disturbances
Beyond the main demonstrations.
Task sequences at a glance
Cross-embodiment validation
The framework is also evaluated on a wheeled mobile manipulation platform.
An additional manipulation task
04 / Reference
Cite this work.
If FALCON is useful to your research, please consider citing our paper.
BIBTEX
arXiv:2512.04381 · 2025
@article{he2025falcon,
title={FALCON: Actively Decoupled Visuomotor Policies for Loco-Manipulation with Foundation-Model-Based Coordination},
author={He, Chengyang and Sun, Ge and Bai, Yue and Lu, Junkai and Zhao, Jiadong and Sartoretti, Guillaume},
journal={arXiv preprint arXiv:2512.04381},
year={2025}
}