RoboMatch: Unified Mobile Manipulation

• RoboMatch is a unified mobile manipulation system that blends cockpit-style teleoperation, imitation learning, and long-horizon execution for enhanced performance.
• It features a synchronized hardware interface that improves demonstration quality by coordinating base and arm motions with virtual reality feedback.
• The system employs a Proprioceptive-Visual Enhanced Diffusion Policy and an Auto-Matching Network to decompose tasks and boost success rates in dynamic environments.

Searching arXiv for the RoboMatch paper and closely related references mentioned in the provided data. RoboMatch is a unified system for mobile manipulation that combines a teleoperation platform, an imitation-learning policy, and a long-horizon execution architecture for long-horizon tasks in dynamic environments (Liu et al., 10 Sep 2025). Its central claim is that mobile manipulation degrades for three coupled reasons: collecting good demonstrations is difficult when base motion and arm motion are not well synchronized; fine manipulation fails when visual and proprioceptive perception are too weak or too coarse; and long-horizon execution degrades when one monolithic policy must handle many heterogeneous subskills over thousands of time steps, causing inference latency and error accumulation. RoboMatch addresses these issues with a cockpit-style mobile-manipulation teleoperation setup for high-quality demonstration collection, a Proprioceptive-Visual Enhanced Diffusion Policy (PVE-DP) for precise short-horizon control, and an Auto-Matching Network (AMN) that decomposes long tasks and routes each subtask to a lightweight specialized policy (Liu et al., 10 Sep 2025).

1. System definition and scope

RoboMatch is presented as a unified whole-body teleoperation and learning system for mobile manipulation rather than as a single policy model (Liu et al., 10 Sep 2025). At the system level, it consists of two physical parts: the robot body and the control cockpit. The robot body includes an indoor differential-drive mobile base (Tracer-2.0), two 7-DoF ViperX-300 slave manipulators, three Logitech C922x RGB cameras, and two IMUs mounted at the end effectors. The cockpit includes two 7-DoF WidowX-250 master arms, foot pedals for base control, a display, keyboard and mouse, an industrial computer, and a Meta Quest 3 VR headset for immersive multi-view monitoring (Liu et al., 10 Sep 2025).

The architecture is intentionally described as “unified”: one operator controls both locomotion and bimanual manipulation, rather than splitting control across two users or loosely coupling separate interfaces (Liu et al., 10 Sep 2025). Demonstrations collected on this platform are used to train short-horizon visuomotor policies. For long-horizon tasks, a VLM-driven planner decomposes a user instruction into sequential subtasks, and AMN dispatches each step to the corresponding pre-trained lightweight policy, with PVE-DP serving as the manipulation policy class (Liu et al., 10 Sep 2025).

This framing places RoboMatch within mobile manipulation, imitation learning, visuomotor policy learning, and long-horizon task decomposition. A plausible implication is that the system should be understood less as a standalone algorithm and more as an integrated research stack spanning demonstration collection, multimodal policy learning, and execution-time orchestration.

2. Cockpit-style teleoperation and demonstration collection

The cockpit-style control interface is the hardware and interaction core of RoboMatch (Liu et al., 10 Sep 2025). Its main idea is an inverted master-slave manipulator layout that behaves like an exoskeleton, so the operator’s arm motions map naturally to robot arm motions at a compatible scale. The dual master arms are used for upper-limb bimanual manipulation, while the pedals control lower-limb mobility. Each pedal contains an IMU; the pedal’s pitch rotation is converted to velocity commands. The left pedal controls forward/backward translation, with speed proportional to pedal tilt. The right pedal controls left/right rotational motion of the base using the same proportional mapping (Liu et al., 10 Sep 2025).

This arrangement creates synchronized upper-lower-limb operation: the operator manipulates the arms while simultaneously driving the base, and watches the scene through VR-assisted multi-view feedback from the robot’s cameras (Liu et al., 10 Sep 2025). The synchronization mechanism is embodied and continuous rather than software-switched: master-arm joint mapping governs slave-arm motion, and pedal IMU pitch controls base velocity, all observed through live visual feedback.

The workflow is direct. The operator wears the VR headset, sees the robot’s multi-view environment, uses the two master arms to puppeteer the dual slave manipulators, and uses the pedals to move the mobile base (Liu et al., 10 Sep 2025). The paper argues that this improves teleoperation precision and stability because control is ergonomic and integrated, and it improves data collection efficiency because a single operator can generate coordinated whole-body demonstrations without handoffs or synchronization lag between separate users.

The data collection pipeline is organized around synchronized multimodal trajectories gathered on both simulated and real platforms (Liu et al., 10 Sep 2025). The recorded data contain visual observations, arm joint states, end-effector quaternion data from IMUs, and actions. The proprioceptive state explicitly includes arm joint positions $X_{\text{qpos}}$ and end-effector orientation quaternions $X_{\text{quat}}$. Visual observations come from RGB cameras, with the paper emphasizing three robot cameras and multi-view visual feedback to the operator (Liu et al., 10 Sep 2025). The paper also describes a step-wise saving strategy for long tasks: when each subtask finishes, the state and action at that boundary are saved as a subtask sample.

The quantitative motivation for this design is explicit. Compared to a split-type data collection platform, RoboMatch yields about a 20% average improvement in one-time collection success rate and produces data that later improves inference success by 7% (Liu et al., 10 Sep 2025). The paper also notes that RoboMatch’s advantage grows with task duration, which suggests that synchronization benefits become more pronounced on longer demonstrations.

3. Proprioceptive-Visual Enhanced Diffusion Policy

PVE-DP is RoboMatch’s short-horizon manipulation policy. It is motivated by observed weaknesses of standard diffusion-policy baselines in fine manipulation, where cumulative inference errors, insufficiently expressive visual features, and limited proprioception hurt delicate tasks such as assembly, grasping, and cloth-like manipulation (Liu et al., 10 Sep 2025). Standard diffusion policies are described as often relying on temporal image features and joint angles alone. RoboMatch augments both streams: on the vision side with FE-EMA, a frequency-enhanced visual module using discrete wavelet transforms, and on the proprioception side with wrist or end-effector IMU orientation sensing fused with joint positions (Liu et al., 10 Sep 2025).

The proprioceptive branch starts from end-effector quaternions measured by IMUs. The paper gives the quaternion representation as

$\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$

with the unit constraint

$$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$

where $w,x,y,z$ are the quaternion components, and $\vec{i},\vec{j},\vec{k}$ are the quaternion basis units (Liu et al., 10 Sep 2025). The fused proprioceptive vector is formed by concatenating joint positions and quaternion states: $$P_{\text{fused}} = \text{Concat}(X_{\text{qpos}}, X_{\text{quat}})$$ This richer self-state representation is then used as conditioning input to the diffusion policy (Liu et al., 10 Sep 2025).

The visual branch introduces a Frequency Enhanced EMA module. The paper applies a Discrete Wavelet Transform (DWT) to feature maps and writes the most general wavelet expansion as

$$X = \sum_{j \in \mathbb{Z}} \sum_{k \in \mathbb{Z}} C_A^{j,k} \cdot \varphi_{j,k}(x) + \sum_{i=1}^{3} C_D^{i,j,k} \cdot \psi_{i,j,k}(x)$$

where $X$ is the input feature map, $j$ is the scale parameter, $X_{\text{quat}}$0 is the translation parameter, $X_{\text{quat}}$1 are approximation coefficients, and $X_{\text{quat}}$2 are detail coefficients (Liu et al., 10 Sep 2025). In the 2D implementation,

$X_{\text{quat}}$3

where $X_{\text{quat}}$4 is the low-frequency approximation and $X_{\text{quat}}$5, $X_{\text{quat}}$6, and $X_{\text{quat}}$7 are horizontal, vertical, and diagonal detail coefficients. The paper uses Haar wavelets. High-frequency features are fused as

$X_{\text{quat}}$8

and combined with the low-frequency approximation through

$X_{\text{quat}}$9

A temporal or spatial attention branch produces $\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$0, and the two streams are fused through

$\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$1

where $\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$2 is a fusion weight balancing temporal-domain features and frequency-domain features (Liu et al., 10 Sep 2025).

The multimodal conditioning then uses both $\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$3 and $\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$4 inside a diffusion policy with a U-Net backbone (Liu et al., 10 Sep 2025). The training objective is written as

$\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$5

where $\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$6 are policy parameters, $\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$7 is the diffusion timestep, $\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$8 is the clean action sequence over horizon $\mathbf{q} = w + x \vec{i} + y \vec{j} + z \vec{k}$9, $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$0 is the sampled Gaussian noise, $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$1 is the noisy version of the action sequence at timestep $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$2, and $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$3 is the network prediction of that noise conditioned on proprioception, vision, noisy actions, and timestep (Liu et al., 10 Sep 2025).

This design locates RoboMatch in the diffusion-policy tradition rather than in flow-matching-based trajectory generation. A relevant comparison point is “Affordance-based Robot Manipulation with Flow Matching,” which represents trajectories as a conditional process of flowing random waypoints to desired robot action trajectories (Zhang et al., 2024). RoboMatch instead retains a diffusion-policy formulation and concentrates its modifications on multimodal conditioning and feature quality rather than replacing the generative paradigm.

4. Auto-Matching Network for long-horizon execution

The Auto-Matching Network addresses the long-horizon problem by replacing a monolithic long-horizon policy with a decomposition-and-routing architecture (Liu et al., 10 Sep 2025). The paper argues that when a single policy must act over 3000–4000 time steps with different perception-action modes across phases of a task, errors accumulate and final success collapses. AMN instead uses a VLM, specifically GLM-4.1V, to parse the current scene and user instruction, generate a chain-of-thought decomposition, and express the task as a sequence of subtasks: $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$4 Here $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$5 is the overall long-horizon task, $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$6 is the $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$7-th subtask, and $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$8 is the specialized policy network assigned to that subtask, parameterized by $$\left|\mathbf{q}\right|^2 = w^2 + x^2 + y^2 + z^2 = 1$$9 (Liu et al., 10 Sep 2025).

During training, demonstrations longer than 3000 time steps are segmented with a step-wise saving mechanism: after each subtask $w,x,y,z$0 completes, the state $w,x,y,z$1 and action $w,x,y,z$2 are saved as subtask samples (Liu et al., 10 Sep 2025). Each short-horizon subtask dataset is then used to train an independent network $w,x,y,z$3. During inference, a user gives a voice or text command; the robot captures the scene image; the VLM combines language and visual context, performs chain-of-thought reasoning, decomposes the task into steps, matches each step to a pre-trained specialized policy, executes that policy, and after each subtask automatically selects the next policy until completion (Liu et al., 10 Sep 2025).

The matching criteria and scheduling logic are described qualitatively rather than as a learned continuous gating network with explicit routing probabilities (Liu et al., 10 Sep 2025). The “matching” is therefore semantic and context-dependent. The VLM reasons over language instructions and the observed visual scene to determine the current subtask and the corresponding lightweight controller. The execution is distributed because different compact policies are run independently for different stages rather than a single massive policy being evaluated over the whole horizon.

This makes RoboMatch’s use of the term “Auto-Matching” specific: it denotes automatic alignment between decomposed task stages and pre-trained specialized policies, not object retrieval or geometric matching. That differs from the usage in “MATCH POLICY,” where matching refers to point cloud registration against demonstrated geometric arrangements to infer pick and place transforms (Huang et al., 2024), and from “RHyME,” where matching refers to sequence-level optimal-transport alignment between human and robot demonstrations under mismatched execution (Kedia et al., 2024). RoboMatch’s “matching” is therefore best understood as subtask-to-policy routing within a long-horizon mobile-manipulation stack.

5. Experimental profile and quantitative results

The paper reports results in simulation and the real world across short-horizon and long-horizon settings (Liu et al., 10 Sep 2025). The short-horizon tasks are Transfer Cube, Insertion Scripted, Wipe Table, and Sort Workpiece. The long-horizon tasks are Clean Trash and Deliver Tool. Clean Trash is a four-step task spanning 3000 time steps, and Deliver Tool is a six-step task spanning 4000 time steps (Liu et al., 10 Sep 2025).

The dataset statistics are explicit. For simulation, the paper uses 100 demonstrations each for Transfer Cube and Insertion Scripted, each totaling 0.2 hours. For real-world short-horizon tasks, it collects 50 demonstrations for Wipe Table (0.3 hours) and 50 demonstrations for Sort Workpiece (0.4 hours). For long-horizon real tasks, it collects 50 demonstrations for Clean Trash (1.1 hours) and 50 demonstrations for Deliver Tool (1.4 hours) (Liu et al., 10 Sep 2025).

The teleoperation platform is evaluated against a split-type baseline platform. RoboMatch improves one-time collection success by about 20% on average relative to that baseline, and data collected with RoboMatch lead to a further 7% higher inference success when training a unified policy (Liu et al., 10 Sep 2025).

For short-horizon learning, Table 5 reports average success rates across four tasks for ACT, DP U-Net, NL-ACT, InterACT, DiT-Block, DP w. EMA, ACT w. quat., DP w. quat., DP w. FE-EMA, and PVE-DP (Liu et al., 10 Sep 2025). The reported averages are:

Method	Average success rate
ACT	52.0
DP U-Net	43.8
NL-ACT	53.8
InterACT	58.8
DiT-Block	56.8
DP w. EMA	54.0
ACT w. quat.	63.0
DP w. quat.	61.0
DP w. FE-EMA	59.0
PVE-DP	75.3

PVE-DP reaches 100, 79, 63, and 59 on the four tasks respectively (Liu et al., 10 Sep 2025). Relative to baseline DP U-Net, the average success rate improvement is $w,x,y,z$4 percentage points (Liu et al., 10 Sep 2025). The paper summarizes this in the introduction and abstract as a 30% success-rate improvement in simulation and 20% in real-world ALOHA-style tasks.

The ablation on parameter and FLOP overhead reports that baseline Diffusion Policy has 101.7214M parameters and 33.9266B FLOPs; DP + FE-EMA has 101.7269M parameters and 33.9662B FLOPs; and DP + FE-EMA + Quat has 101.8416M parameters and 33.9664B FLOPs (Liu et al., 10 Sep 2025). This suggests that the reported gains are not attributed to large increases in model size.

For long-horizon execution, the paper reports final completion rates for AMN and non-AMN baselines (Liu et al., 10 Sep 2025). In Clean Trash, final completion at the “Back” stage is 50% for AMN versus 10% for ACT, 0% for DP, and 8% for PVE-DP. In Deliver Tool, final completion is 44% for AMN versus 2% for ACT and 0% for both DP and PVE-DP (Liu et al., 10 Sep 2025). Intermediate-stage numbers further show AMN’s stability: for Clean Trash it achieves 74/70/54/50% over Grasp/Move/Throw/Back, while PVE-DP gets 66/42/20/8; for Deliver Tool it achieves 100/68/66/62/48/44 over the six stages, while PVE-DP gets 100/54/28/8/0/0 (Liu et al., 10 Sep 2025).

These results support three distinct empirical claims: improved data collection efficiency through the cockpit, improved fine manipulation through PVE-DP, and improved long-horizon execution through AMN (Liu et al., 10 Sep 2025). A plausible implication is that RoboMatch’s contribution is cumulative: each subsystem addresses a separate failure mode, and the reported gains are distributed across collection, control, and execution rather than concentrated in one module alone.

6. Relation to adjacent methods, limitations, and significance

RoboMatch belongs to a broader landscape of robotic systems that use “matching” in different technical senses. “MATCH POLICY” reformulates high-precision manipulation as point cloud registration against stored demonstrations and derives actions analytically from transforms such as

$w,x,y,z$5

(Huang et al., 2024). “RHyME” performs sequence-level optimal-transport retrieval to compose semantically equivalent human demonstrations for robot policies under cross-embodiment execution mismatch (Kedia et al., 2024). “RoboEye” treats robotic object identification as query-to-gallery retrieval and introduces selective 3D reranking with confidence-weighted keypoint correspondence matching for warehouse retrieval (Zhang et al., 18 Sep 2025). RoboMatch differs from all three: its matching mechanism is policy routing for long-horizon mobile manipulation rather than registration, retrieval, or demonstration alignment.

On the policy side, RoboMatch is closer to diffusion-policy work than to flow-matching work. “Affordance-based Robot Manipulation with Flow Matching” replaces diffusion with deterministic ODE-based trajectory generation and reports faster inference with competitive or slightly better trajectory accuracy (Zhang et al., 2024). RoboMatch instead retains a diffusion policy and adds FE-EMA plus quaternion-enhanced proprioception (Liu et al., 10 Sep 2025). This suggests that its design priority is multimodal sensory enrichment within a diffusion framework rather than a change of generative backbone.

The paper mentions limitations and future work briefly but clearly (Liu et al., 10 Sep 2025). The cockpit is slated for further optimization for operator comfort and convenience during data collection. PVE-DP is to be refined further for better manipulation performance. For AMN, the authors explicitly mention reducing the “gaps between subtasks,” indicating a present weakness at subtask boundaries. More broadly, the method assumes that long tasks can be meaningfully decomposed into reusable subtasks and that suitable pre-trained policies exist for each one (Liu et al., 10 Sep 2025). The AMN routing is also VLM-dependent, so planning or perception errors at the semantic layer can still cause failures. The paper does not deeply discuss ambiguous instructions, unseen subtasks, severe visual occlusion, or dynamic perturbations beyond its general claim of robustness in dynamic environments.

The significance of RoboMatch lies in the way it ties together whole-body teleoperation, multimodal short-horizon policy learning, and semantic long-horizon orchestration within a single mobile-manipulation stack (Liu et al., 10 Sep 2025). Rather than proposing one monolithic universal controller, it treats synchronized demonstration collection, enriched perceptual-proprioceptive conditioning, and subtask-level model dispatch as separate engineering and learning problems. This suggests a systems view of mobile manipulation in which long-horizon competence emerges from coordinated design across interfaces, data, policy architecture, and execution-time decomposition rather than from scale alone.