MobileVLA-R1: Reinforcing Vision-Language-Action for Mobile Robots¶

Conference: ECCV2026
Paper: ECCV Paper
Project: MobileVLA-R1
Area: Robotics & Embodied AI
Keywords: vision-language-action models, multi-granularity chain-of-thought, offline GRPO, quadruped robots, closed-loop control

TL;DR¶

MobileVLA-R1 establishes a parseable reasoning-and-action interface through multi-granularity chain-of-thought supervision, then uses offline GRPO rewards for continuous movement direction, discrete behaviors, and output format, reaching 68.3% SR on R2R-CE Val-Unseen and demonstrating quadruped closed-loop execution through fixed onboard/remote hybrid deployment.

Background & Motivation¶

Understanding "go around the box, approach the door, and lie down" does not tell a robot how fast to move now, which way to turn, or when to change posture. A vision-language model can identify objects and describe a scene, but a quadruped ultimately needs continuous velocities, yaw rates, and concrete behavior commands. Under partial observability, each action changes the next observation, so an early incorrect turn can prevent later language goals from being grounded. Vision-language-action models such as NaVILA already connect language navigation with legged motion; this paper focuses on consistency between explicit reasoning and executable outputs at that interface.

Direct action imitation can learn local control without revealing whether the model distinguishes targets, obstacles, and the ordering of subtasks. Conversely, a natural-language plan is insufficient: it may sound correct while the final answer is unparseable or its movement direction contradicts the plan. The authors therefore treat chain-of-thought (CoT) as more than an additional explanation, requiring one output to contain both rationale text and a structured command so training feedback reaches the fields sent to execution. The key constraint is command verifiability, not having another language model judge prose quality or performing online reinforcement learning on the physical robot.

The paper organizes trajectory strategies, current-step decisions, and long-horizon navigation into different supervision granularities, establishing a stable interface through supervised fine-tuning before comparing multiple candidate outputs within groups. Training uses existing observations, histories, and command targets; only deployment obtains subsequent observations from the real environment. Core Idea: first learn to express multi-scale task understanding as executable structured answers, then use offline command-level rewards to reinforce what is actually output after reasoning, rather than merely making explanations sound plausible.

Method¶

Overall Architecture¶

Training comprises Multi-granularity CoT Data, Reasoning-Execution Alignment, and Offline GRPO; deployment consists of observation input, structured output, deterministic parsing, and closed-loop execution. The model receives RGB, depth, point clouds, and language instructions, while data samples can also include task-specific states and action histories. Its outputs are not joint torques: they are planar velocities, yaw rates, and discrete high-level behaviors such as posture or skill switching. This interface preserves the responsibility of the existing low-level controller rather than presenting high-level language generation as direct control of every motor.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    Source["R2R / RxR / QUARD"] --> Data["Multi-granularity<br/>CoT Data"]
    Data --> Align["Reasoning-Execution<br/>Alignment"]
    Align --> RL["Offline GRPO"]
    RL -->|Trained policy| Model["Structured rationale<br/>and answer"]
    Obs["Current observations<br/>and instruction"] --> Model
    Model --> Decode["Deterministic parsing<br/>and low-level execution"]
    Decode -->|Next real observation| Obs

Candidate sampling and reward optimization occur during training, not as a GRPO update every time the robot takes a step. The deployed policy regenerates an answer from new observations, but its parameters do not adapt online. The paper describes this as offline reasoning-to-command consistency optimization, with its effects subsequently evaluated in closed-loop tasks.

Key Designs¶

1. Multi-granularity CoT Data: connecting long-horizon intent and immediate control through one supervision interface

MobileVLA-CoT is constructed from R2R, RxR, and QUARD: the first two supply language/navigation trajectories, while the last adds quadruped locomotion and manipulation observations with control targets. Episode-level supervision summarizes trajectory outcomes and strategies, Step-level supervision explains the next action under the current observation, and Navigation-level supervision connects global instructions with multiple decisions along a route. These are not three agents running concurrently; they are ways of organizing training examples for the same policy. The final subsets contain 18K, 78K, and 38K examples, totaling 134K; source-dataset sizes in Table 1 count different units from these synthesized examples. Trajectory-level information helps preserve task order, while step-level information grounds rationales in concrete control fields, addressing errors at different temporal scales.

The data engine uses Gemini-2.5-Flash to process multimodal observations, instructions, and optional histories, producing rationale and target output with <think> and <answer> tags. The authors describe the prompts, parsing, and verification as backend-independent, although Gemini-2.5-Flash remains the teacher used for the main dataset. Quality control covers format, action consistency, safety, and manual checks, retaining 134K of 168K raw generations. Filtering removes non-executable outputs, incorrect commands, visual inconsistencies, and hallucinations, so the dataset is not an unfiltered collection of model explanations. However, filtering does not establish that every rationale faithfully captures the policy's internal causal decisions; it primarily establishes usable consistency between supervision and action targets.

2. Reasoning-Execution Alignment: first establishing a stable perception, language, and control interface

The model is initialized from NaVILA, organizes multimodal inputs in a LLaVA-style architecture, and uses LLaMA3-8B as its language backbone. RGB, depth, and point clouds are encoded separately, projected into the language hidden dimension through learned linear adapters, and concatenated with text tokens. DepthAnything V2 encodes depth and Point Transformer v3 encodes point clouds; modality-type embeddings distinguish token sources, although the implementation description lists them as optional. This fusion introduces geometry directly into the language decision context without first converting a point cloud into a spatially lossy textual description. Per-modality token counts control the input budget, but the inspected main text does not specify these counts, so a fixed inference cost cannot be claimed.

Cold-start training first uses Episode and Navigation supervision for trajectory understanding and structured reasoning, then continues with executable Step answers to align control. Each output follows <think>...</think><answer>...</answer>, and a deterministic parser converts the answer into velocities and discrete behaviors. The action decoder in the described method is therefore primarily a structured-command parsing and execution interface, not an unspecified diffusion action head or newly trained joint controller. This SFT initialization reduces the risk of starting reinforcement learning with format drift, unparseable text, and invalid behavior fields. All modality encoders remain frozen, while LoRA adapts multimodal projection layers and attention modules in the language backbone, concentrating training on cross-modal alignment and output behavior.

3. Offline GRPO: rewarding parsed commands rather than only comparing text similarity

For the same observation and instruction, the policy samples 8 complete candidate outputs and parses each into continuous movement quantities and a discrete behavior. The movement reward first divides translational velocities by \(s_v\) and yaw rate by \(s_\omega\), avoiding a direct comparison of quantities with different units. After obtaining the normalized predicted command vector \(\hat{\mathbf{u}}\) and target vector \(\mathbf{u}^{*}\), the reward measures their directional agreement. The following is equivalent notation for the mechanism described by Equations (6)-(7) on page 8; the cached typesetting is broken, and the other damaged equations are not reconstructed here.

\[ R_m = \frac{\hat{\mathbf{u}}^{\top}\mathbf{u}^{*}}{\lVert\hat{\mathbf{u}}\rVert\lVert\mathbf{u}^{*}\rVert}. \]

This expression requires nonzero vector norms; the inspected main text does not specify special handling for zero-speed commands or numerical values for the normalization scales. Cosine reward emphasizes direction rather than absolute speed magnitude, so a high reward alone does not establish accurate speed or low physical collision risk. The action reward is 1 when the discrete behavior exactly matches its target and 0 otherwise; the format reward is 1 when the output follows the required rationale-and-answer template and 0 otherwise. The three rewards are combined with fixed weights, constraining continuous direction, discrete behavior, and machine parseability respectively. They are derived from offline command targets and output format, not directly from whether an actual rollout ultimately succeeds.

GRPO subtracts the group mean reward from each candidate's reward and normalizes by the within-group standard deviation, favoring relatively better answers to the same input. Updates use a clipped policy probability ratio and KL regularization to prevent excessive deviation from the frozen SFT reference policy. This differs from multiplying supervised cross-entropy by a sample weight: the policy generates complete candidates and then optimizes their output probabilities based on the resulting commands. The trade-off is dependence on reward coverage: a parseable, directionally correct command does not automatically imply correct reasoning or long-horizon safety.

A Worked Example¶

Consider the instruction in Figure 4 to pass through the gap between a sofa and a chair and then sit down, which requires movement before a posture change. RGB supplies object identity, depth and point clouds provide gap geometry, and history can help establish whether the target region has already been traversed. The model first generates a rationale for its current decision, then supplies movement velocity, turning, and behavior fields that the parser passes to low-level control. New observations after movement return to the policy, and subsequent answers determine whether to continue or sit, rather than generating one long plan and executing it blindly in open loop. During training, 8 candidates for the same input receive different rewards for direction, behavior matching, and format validity; deployment has no such scoring loop with target answers. This example explains the mechanism and does not imply that the paper publishes an additional successful trajectory with complete stepwise velocity values.

Loss & Training¶

SFT uses AdamW and a cosine learning-rate schedule at \(2\times10^{-4}\) on 4 H20 GPUs with 96GB each. LoRA uses rank 16 and scaling 32; the sequence of long-horizon reasoning/schema initialization followed by stepwise control alignment comes from Section 4.2. GRPO uses one H20 with 96GB for 1K optimizer steps, with a default batch of 5 inputs per update and 8 candidates per input. Its AdamW learning rate is \(1\times10^{-6}\), the KL coefficient is 0.04, the clipping coefficient is 0.2, and the SFT reference policy stays frozen. Table 7 sweeps movement-reward weights of 0.00, 0.10, 0.25, 0.50, and 1.00, but the main text does not clearly list absolute values for the other two reward weights. These missing parameters and zero-norm handling affect reproducibility and should not be replaced with assumed GRPO defaults.

Key Experimental Results¶

Main Results¶

The following selection is from Table 2 on page 9, evaluated on VLN-CE Val-Unseen; SR measures success and SPL combines success with path efficiency, both higher-is-better and reported on a percentage scale. NaVILA and CorrectNav use single-view RGB without depth; MobileVLA-R1 is marked as single-view RGB plus depth, and all three lack panoramic observations and odometry. This is therefore not a strictly sensor-matched algorithm ablation; the paper also discusses point-cloud encoding, but Table 2 has no separate point-cloud input column.

Dataset	Method	SR ↑	SPL ↑
R2R-CE Val-Unseen	NaVILA	54.0	49.0
R2R-CE Val-Unseen	CorrectNav	65.1	62.3
R2R-CE Val-Unseen	MobileVLA-R1	68.3	65.2
RxR-CE Val-Unseen	NaVILA	49.3	44.0
RxR-CE Val-Unseen	CorrectNav	69.3	63.3
RxR-CE Val-Unseen	MobileVLA-R1	71.5	66.8

Relative to CorrectNav, SR improves by 3.2 percentage points on R2R-CE and 2.2 on RxR-CE; the abstract's "about 5%" must not be applied indiscriminately to every comparison. In the real-world Table 4 on page 12, complex Corridor SR increases from NaVILA's 0.67 to 0.86, while NE decreases from 1.76 to 1.23. Real-world evaluation comprises 160 episodes; success requires reaching the target, completing the requested behavior, and no human intervention, while Outdoor reports only SR because global localization is unreliable. QUARD Table 3 on page 10 reports average SR of 0.73 for MobileVLA-R1 and 0.60 for MoRE, but the arithmetic mean of the six listed MobileVLA-R1 task values is approximately 0.698, inconsistent with the stated average and retained here as a source uncertainty.

Ablation Study¶

Table 5 on page 12 ablates rewards on R2R-CE Val-Unseen, with SR and SPL on a percentage scale; the text identifies the no-reward row as the SFT-only baseline.

Movement Reward	Action Reward	Format Reward	SR ↑	SPL ↑
Disabled	Disabled	Disabled	58.0	53.2
Enabled	Disabled	Disabled	60.7	55.5
Disabled	Enabled	Disabled	61.9	56.8
Disabled	Disabled	Enabled	59.6	54.7
Enabled	Enabled	Disabled	64.5	60.2
Enabled	Disabled	Enabled	63.4	59.1
Disabled	Enabled	Enabled	65.2	61.0
Enabled	Enabled	Enabled	68.3	65.2

The full reward combination gains 10.3 SR points and 12.0 SPL points over SFT-only; removing the action reward reduces SR to 63.4, the largest drop among single-component removals. Table 8 on page 14 controls the teacher, prompt, output format, and rationale-token budget per episode while comparing CoT granularities; No-CoT retains action targets and command format but removes rationale text.

CoT Supervision	SR ↑	SPL ↑	SR Gain over No-CoT
No-CoT, actions only	64.0	59.6	+0.0
Episode-only CoT	65.5	61.3	+1.5
Step-only CoT	66.0	61.7	+2.0
Navigation-level CoT	66.7	62.6	+2.7
Multi-granularity	68.3	65.2	+4.3

No-CoT here is not the preceding table's SFT-only condition: one removes rationale supervision, while the other removes reinforcement-learning rewards, so their difference cannot directly quantify the benefit of CoT.

Key Findings¶

Combined rewards outperform any single component, but the ablation alone does not establish that natural-language rationales faithfully reflect internal decisions.
Multi-granularity supervision gains 4.3 SR points over No-CoT in Table 8, a more specific result than claiming that adding any rationale text helps.
Under matched rewards, schema, KL regularization, updates, and sampling budget, Table 9 on page 14 reports 68.3 SR / 65.2 SPL for GRPO versus 64.1 / 60.4 for PPO.
Reward-SFT reaches 62.4 SR / 58.3 SPL in Table 11 on page 15, indicating that supervised sample reweighting does not fully account for the command-reward benefit in this setting.

Highlights & Insights¶

Placing rationales and commands in one parseable output lets action-level errors feed back into the generation policy, bringing the method closer to a robot execution interface than displaying a separate explanation.
Multi-granularity supervision divides attention between long-horizon task order and immediate control; other tasks can reuse trajectory intent plus stepwise executable labels without copying these particular datasets.
Format reward is not intelligence by itself, but it determines whether language outputs can enter a control loop; only its joint optimization with behavior and movement direction forms the full alignment objective here.

Limitations & Future Work¶

The authors explicitly acknowledge remote-inference dependence: perception, mapping, parsing, and control run onboard, while the 8B forward pass is always offloaded to an H20; this is neither fully onboard nor dynamic local/remote switching.
Reported end-to-end latency is 205-245 ms and is described as compatible with 5 Hz high-level control; a 5 Hz period is 200 ms, however, and the text does not explain asynchronous pipelining or deadline handling, so strict real-time guarantees are not established.
QUARD averages have the inconsistency noted above; some task success rates also do not follow the 0.04 increment of a single 25-episode count, and the inspected main text does not explain further averaging, so the original table is not silently corrected.
Main comparisons do not fully match sensor inputs and the inspected tables lack error bars; attributing total gains to geometry versus training requires further controlled evaluation.
Some formulas and figure text are corrupted in the cache, and supplementary safety-filtering and failure-breakdown details are outside the inspected material; zero-speed rewards and execution fallback remain to be verified.
Future work could explore speed-magnitude and collision-constraint rewards, safe fallback during disconnection, and onboard compression; these are reader suggestions, not completed experiments in the paper.

NaVILA: the paper directly inherits its embodied-alignment initialization, then adds geometry, multi-granularity CoT, and offline command rewards, making the contribution primarily an extension and post-training strategy rather than a legged controller built from scratch.
QUART / QUAR-VLA: these also address quadruped vision-language-action tasks; MobileVLA-R1 additionally connects explicit reasoning structure with parseable continuous commands.
PPO and Reward-SFT: the former provides a policy-optimization comparison, while the latter tests whether rewards merely select better supervised examples; matched-budget comparisons more directly support the role of group-relative optimization.
Research implication: testing whether CoT causally determines actions could involve intervening on rationale content while holding the command target and scene fixed, then measuring behavior changes; this is further causal validation, not a conclusion already established by the success-rate tables.

Rating¶

Novelty: 4/5. Combines multi-granularity reasoning supervision, structured commands, and offline GRPO at the quadruped execution interface, with novelty primarily in their integration.
Experimental Thoroughness: 4/5. Covers navigation, quadruped tasks, real robots, and multiple controlled ablations, although input differences and statistical uncertainties limit the strength of conclusions.
Writing Quality: 3/5. Clearly distinguishes training from deployment, but some statistics and real-time claims need clarification, and cached equation formatting impedes reading.
Value: 4/5. Offers a concrete recipe for executable post-training of reasoning-oriented VLA models, while remaining short of fully onboard, strictly safe real-time operation.