PaperlayerView exact revision on arXiv

Source-linked AI summary

MobileVLA-R1 2.0: RL-Enhanced Reasoning for Mobile Robot Control

Ting Huang, Yue Huang, Zeyu Zhang, Shuicheng Yan, Hao Tang

arXiv:2609.06251v1cs.CVcs.RO

TL;DR

MobileVLA-R1 2.0 addresses the gap between semantic reasoning and executable mobile-robot control. It combines multi-granularity CoT supervision and reinforcement learning with a reasoning-conditioned decoder for task-level actions. It reports gains in navigation and real-world G1 mobile manipulation over MobileVLA-R1.

  • Problem

    Mobile-robot VLA systems still face a gap between high-level reasoning and executable control, especially when tasks require coordinated locomotion and manipulation.

    Source: §1 · pp. 1–2

  • Method

    MobileVLA-R1 2.0 trains multi-granularity reasoning with supervised CoT alignment and GRPO, then decodes reasoning into task-level actions realized by robot-specific controllers.

    Source: §1 · p. 2, §4.1 · p. 5

  • Results

    MobileVLA-R1 2.0 improves VLN-CE SR by 1.6 points and real-world G1 full-task success by 10.0 points over MobileVLA-R1.

    Source: §1 · p. 3

  • Takeaways & Limitations

    G1 evaluation without embodiment-specific policy fine-tuning supports transfer of the learned reasoning-to-action capability to humanoid mobile manipulation.

    Source: §1 · p. 3, §8 · p. 16

    The real-world system runs its 8B VLA backbone on a remote GPU because of onboard memory constraints, adding communication overhead and network dependence.

    Source: §7 · p. 16

Abstract

from arXiv · show

Grounding natural-language instructions into reliable and executable actions remains a fundamental challenge for vision-language-action (VLA) systems on mobile robots, due to the persistent gap between high-level semantic reasoning and low-level locomotion and manipulation control. Existing approaches often rely on implicit reasoning or monolithic action prediction, making it difficult to maintain coherent long-horizon decision making while producing precise and adaptable robot actions. To address this challenge, we propose MobileVLA-R1 2.0, an RL-enhanced VLA framework that explicitly couples structured embodied reasoning with executable mobile robot control. The framework learns multi-granularity reasoning over embodied trajectories through supervised Chain-of-Thought (CoT) alignment and reinforcement learning, improving reasoning-to-action consistency beyond purely behavioral supervision. To support both locomotion and manipulation, we further introduce a reasoning-conditioned action decoder that maps multimodal reasoning representations to task-level action targets, which are subsequently translated into embodiment-specific commands by robot controllers. This design provides a unified perception-reasoning-action interface while decoupling high-level action generation from robot-specific actuation. We conduct extensive evaluations on language-guided navigation, quadruped control, and humanoid mobile manipulation, covering VLN-CE, QUARD, and real-world deployments on Unitree Go2 and G1 robots. MobileVLA-R1 2.0 consistently outperforms strong VLA baselines, achieving an average 1.6 point improvement in SR on VLN-CE and a 10.0 point improvement in full-task success on real-world G1 mobile manipulation tasks over MobileVLA-R1, while demonstrating robust long-horizon instruction following and closed-loop execution across different robotic platforms.

1 INTRODUCTION

Mobile robot control requires grounding semantic decisions in executable behavior across partial observability, long-horizon tasks, locomotion, and manipulation. MobileVLA-R1 2.0 couples structured reasoning to task-level actions and robot-specific control, with reported gains on navigation and humanoid manipulation.

1 INTRODUCTION

  • Reliable mobile-robot control must ground language decisions in physical actions amid partial observability, sensing uncertainty, actuation noise, and long-horizon dependencies.Mobile manipulation additionally requires coordinated locomotion and physical interaction.Source: §1 · p. 1
  • Existing VLA policies often leave intermediate reasoning implicit, while embodied-reasoning methods leave its connection to executable heterogeneous-robot control underexplored.The challenge includes converting deliberative representations into compact actions that heterogeneous controllers can realize.Source: §1 · pp. 1–2
  • The framework combines multi-granularity embodied reasoning, supervised CoT alignment, and reinforcement learning with a reasoning-conditioned action decoder.The decoder produces task-level locomotion and behavior targets, which robot-specific low-level controllers translate into executable commands.Source: §1 · pp. 2–3
  • MobileVLA-R1 2.0 is evaluated on VLN-CE, QUARD, and real-world Unitree Go2 and G1 deployments.The G1 experiments assess object search, navigation, target approach, grasping, and long-horizon combinations of these skills.Source: §1 · pp. 2–3
  • MobileVLA-R1 2.0 improves VLN-CE SR by 1.6 points and real-world G1 full-task success by 10.0 points over MobileVLA-R1.The G1 evaluation includes long-horizon mobile-manipulation tasks and uses no G1-specific trajectories or task annotations during training.Source: §1 · p. 3

2 RELATED WORK

Prior work advances language-guided navigation, generalist VLA policies, and explicit reasoning, but reliable coupling of structured reasoning with coordinated mobile control remains comparatively underexplored.

Language-guided navigation and quadruped VLA

  • Earlier VLN and quadruped methods advance language-guided mobility, while generalist and reasoning-enhanced VLA work expands action generation and intermediate reasoning,,,,,, [13], [14].The section contrasts these directions with the explicit grounding of structured reasoning into executable continuous control.Source: §2 · pp. 3–4

Generalist and reasoning-enhanced VLA

  • MobileVLA-R1 2.0 combines multi-granularity CoT supervision, GRPO-based optimization, and a reasoning-conditioned decoder to connect structured reasoning with locomotion and manipulation control.The cited literature includes ECoT, CoT-VLA, ACoT-VLA [13], and dense embodied reasoning methods [14].Source: §2 · p. 4

RL-enhanced VLA and mobile manipulation

  • The literature also covers reinforcement learning for quadruped VLA control and vision-language manipulation,, alongside mobile-manipulation and humanoid-control approaches, [16].The table caption identifies statistics for source datasets and MobileVLA-CoT by Nav., Emb-Ctl., CoT, and sample counts.Source: §2 · p. 4, Table 1 · p. 4

3 DATASETS

MobileVLA-CoT is built from navigation and quadruped-control data, then augments trajectories with reasoning at episode, step, and navigation granularity. A model-agnostic engine generates and verifies reasoning traces paired with executable targets.

3.1 Source Datasets

  • R2R, RxR, and QUARD supply complementary navigation and quadruped-control supervision; no G1 trajectories, demonstrations, or task annotations are used for training.Unitree G1 is introduced only during evaluation to assess transfer to humanoid mobile manipulation.Source: §3.1 · p. 4

3.2 Multi-Granularity Embodied Reasoning Dataset

  • MobileVLA-CoT adds episode-, step-, and navigation-level reasoning to trajectories, with 18K, 78K, and 38K samples respectively, totaling 134K annotated instances.Samples pair observations, instructions, optional state–action history, a structured reasoning trace, and an executable target.Source: §3.2 · p. 4

3.3 CoT Data Engine and Quality Control

  • The CoT data engine generates reasoning traces and executable outputs from multimodal observations, instructions, and optional histories; verification reduces 168K raw generations to 134K validated samples.The four-stage checks cover format validity, action consistency, safety, and semantic correctness; Fig. 2 depicts reasoning-conditioned prediction of locomotion and a discrete behavior primitive.Source: §3.3 · pp. 4–5, Figure 2 · p. 5

4 THE PROPOSED METHOD

MobileVLA-R1 2.0 uses a hierarchical reasoning–execution pipeline to map multimodal observations and instructions to task-level actions, with supervised alignment followed by GRPO optimization. The action decoder separates task-level action generation from embodiment-specific execution.

4.1 Overview and Problem Formulation

  • MobileVLA-R1 2.0 generates structured reasoning from multimodal observations and instructions, then maps reasoning representations to task-level actions in a two-stage SFT-then-GRPO training pipeline.Task-level outputs specify intended behavior rather than morphology-specific joint commands, and robot-specific low-level controllers execute them.Source: §4.1 · p. 5

4.2 Multimodal Reasoning Backbone

  • The multimodal backbone encodes RGB, depth, and point-cloud observations with modality-specific encoders and contextualizes their projected tokens with language tokens in a VLA backbone.The contextual observation states and generated reasoning-token states provide perceptual and reasoning context to the action decoder.Source: §4.2 · pp. 5–6

4.3 Reasoning-Conditioned Action Decoder

  • The reasoning-conditioned decoder maps observation and reasoning representations to locomotion and behavior predictions, while robot-specific controllers translate task-level actions into embodiment-specific commands.Locomotion is represented by planar translational and yaw velocities; behavior is selected from a predefined task-level set, with the highest-probability primitive used at inference.Source: §4.3 · p. 6

4.4 Supervised Reasoning and Action Alignment

  • Supervised alignment first uses episode- and navigation-level data for long-horizon reasoning, then step-level data to jointly optimize the action decoder with CoT supervision.The supervised objective links intermediate reasoning to task-level physical actions and activates locomotion or behavior losses only when corresponding targets are available.Source: §4.4 · pp. 6–7

4.5 GRPO-Based Reasoning-to-Action Optimization

  • The method applies offline GRPO after supervised alignment, using group-relative advantages and movement, behavior, and format rewards while KL regularization constrains deviation from a frozen reference policy.The decoder remains fixed during GRPO; the trainable VLA policy is optimized from rewards derived from actions induced by sampled reasoning trajectories.Source: §4.5 · pp. 7–8, Figure 3 · p. 7

5 EXPERIMENTS

The experiments cover VLN-CE, QUARD, and real-world Go2 and G1 deployments, with reported gains over prior models in navigation, quadruped control, and humanoid mobile manipulation.

5.1 Experimental Setup

  • The setup evaluates R2R-CE and RxR-CE val-unseen navigation, six QUARD tasks, and real-world Go2 control and G1 manipulation without G1-specific training data.Navigation uses NE, OS, SR, SPL, and nDTW; QUARD reports task-wise and average success rates.Source: §5.1 · pp. 8–9

5.2 Main Results

  • On R2R-CE and RxR-CE val-unseen splits, MobileVLA-R1 2.0 reaches SRs of 69.8 and 73.1, improving over MobileVLA-R1 by 1.5 and 1.6 percentage points; on QUARD, average success rises from 0.70 to 0.76.Source: §5.2 · p. 10

5.3 Real-World Evaluation

  • On Go2 complex tasks, Workspace SR rises from 0.91 to 0.94 and Corridor SR from 0.86 to 0.91, while NE falls from 1.23 to 1.12 and 1.23 to 1.11, respectively.The report identifies grounding, obstacle handling, excessive turning, narrow passages, and localization drift among observed failures.Source: §5.3 · p. 11
  • On G1, average full-task success rises from 46.7% to 56.7% versus MobileVLA-R1, a 10.0-point gain without G1-specific policy fine-tuning.Conditional manipulation success also increases in Tabletop, Shelf/Cabinet, and Cluttered scenarios.Source: §5.3 · pp. 12–13, Table 6 · p. 13

5.4 Ablation Study

  • Combining learnable action decoding with GRPO performs best, and the combination improves R2R-CE SR from 68.3 to 69.8 and SPL from 65.2 to 66.9 over GRPO alone.The decoder and GRPO each improve over the supervised baseline; GRPO gives the larger individual gain.Source: §5.4 · p. 14, Table 9 · p. 14
  • Combining episode-, navigation-, and step-level reasoning supervision yields R2R-CE SR of 68.3 versus 64.0 and SPL of 65.2 versus 59.6.All single-granularity variants outperform No-CoT, with navigation-level supervision strongest among the individual variants.Source: §5.4 · pp. 14–15, Table 10 · p. 15
  • The full GRPO reward configuration achieves the best R2R-CE SR of 68.3 and SPL of 65.2; behavior reward is strongest among individual reward terms.Movement alignment, behavior correctness, and output-format validity provide complementary reward signals.Source: §5.4 · p. 15, Table 11 · p. 15
  • The Dual-Query Dual-Head decoder reaches 69.8 SR and 66.9 SPL, improving 1.1 and 1.3 points over Mean Pooling + MLP.Attention-based decoding outperforms mean pooling, with further gains from separating locomotion and behavior branches.Source: §5.4 · p. 15, Table 12 · p. 15

6 TEST-TIME EFFICIENCY

The system runs sensing and low-level control onboard while sending the 8B VLA backbone and action decoder to a remote GPU. This hybrid setup has measured end-to-end latency of 205–245 ms, while onboard memory constraints prevent fully local inference.

6 TEST-TIME EFFICIENCY

  • Sensing, mapping, preprocessing, and low-level control run onboard; the 8B VLA backbone and reasoning-conditioned decoder run remotely due to Jetson Orin Nano memory constraints.The remote policy returns task-level commands for robot execution without deterministic parsing of textual action outputs.Source: §6 · p. 15
  • 205–245 ms end-to-end latency corresponds to an effective high-level decision rate of approximately 4.1–4.9 Hz.Latency spans observation acquisition through task-level command handoff; low-level controllers operate independently at higher frequencies.Source: §6 · p. 15
  • The 8B backbone remains the main computational and memory bottleneck, preventing fully onboard inference on the current Jetson platform.Remote inference may add latency and reduce robustness under constrained or unstable network connectivity.Source: §6 · p. 15

7 LIMITATIONS AND FUTURE WORK

MobileVLA-R1 2.0 has limits in action expressiveness, embodiment coverage, and deployment. Future work targets richer control, broader adaptation, and onboard inference.

7 LIMITATIONS AND FUTURE WORK

  • The task-level action space does not directly represent fine-grained contact dynamics, dexterous manipulation, or continuously parameterized whole-body motion.It combines continuous locomotion commands with a finite set of behavior primitives.Source: §7 · p. 15
  • G1 transfer without G1-specific policy fine-tuning does not establish general embodiment adaptation; performance may degrade across different morphologies, sensing configurations, environments, or task distributions.Training uses supervised trajectories and multi-granularity reasoning annotations from the considered navigation and robot-control datasets.Source: §7 · pp. 15–16
  • The real-world system runs its 8B VLA backbone on a remote GPU because of onboard memory constraints, adding communication overhead and dependence on network connectivity.Latency nevertheless supports the closed-loop tasks evaluated in this work.Source: §7 · p. 16
  • Future work proposes richer parameterized skills and hierarchical whole-body control, broader cross-platform training and online adaptation, and compression, quantization, and distillation toward onboard inference.The authors also identify dexterous manipulation, dynamic human–robot interaction, and longer-horizon open-world tasks as directions for extension.Source: §7 · p. 16

8 CONCLUSION

MobileVLA-R1 2.0 connects multi-granularity reasoning to task-level locomotion and behavior prediction. Evaluations show improvements across navigation and control benchmarks and real-world robot deployments, including G1 mobile manipulation without embodiment-specific policy fine-tuning.

8 CONCLUSION

  • The framework combines multi-granularity reasoning supervision, a reasoning-conditioned action decoder, and GRPO-based optimization to connect intermediate reasoning with task-level locomotion and behavior prediction.The authors present it for language-guided mobile robot control.Source: §8 · p. 16
  • Experiments on R2R-CE, RxR-CE, and QUARD demonstrate consistent improvements in navigation and reasoning-aligned control.Real-world evaluations on Unitree Go2 and Unitree G1 further validate closed-loop mobile control and humanoid mobile manipulation, respectively.Source: §8 · p. 16
  • G1 evaluation requires no embodiment-specific policy fine-tuning, and the authors conclude that explicit reasoning-to-action alignment is valuable for interpretable and executable VLA policies for long-horizon embodied tasks.Source: §8 · p. 16

APPENDIX A ADDITIONAL ABLATION STUDY

The ablations examine geometric modalities, movement-reward weighting, rationale sources, and optimization choices. They report gains from richer perception and reasoning supervision, with GRPO and a movement-reward weight of 1.0 strongest among tested settings.

Effect of multimodal perception

  • 68.3 SR and 65.2 SPL on R2R-CE, and 71.5 SR and 66.8 SPL on RxR-CE, are reached by the complete configuration; depth plus point clouds reaches 67.2 SR and 63.5 SPL on R2R-CE and 70.2 SR and 65.2 SPL on RxR-CE.Adding depth to text and RGB improves both SR and SPL, and point clouds provide additional gains; geometric representations complement RGB appearance for spatially grounded navigation.Source: Table 13 · p. 19, Appendix A · p. 19

Sensitivity to the movement reward weight

  • At λmov = 1.0, navigation performance is best among tested settings, with SR increasing from 65.2 to 68.3 and SPL from 61.0 to 65.2 as λmov increases.The normalized contribution of Rmov rises from 0% to 32%, while behavior and format rewards together remain the dominant source of reward-derived advantage; movement alignment is complementary.Source: Table 14 · p. 19, Appendix A · p. 19

Effect of rationale source

  • Gemini-CoT performs best among the tested rationale sources, achieving 68.3 SR and 65.2 SPL.Template-CoT improves over No-CoT and model-generated rationales provide further gains; identical action targets mean the differences primarily reflect rationale supervision.Source: Table 15 · p. 19, Appendix A · p. 19

Effect of policy optimization strategy

  • GRPO improves R2R-CE Val-Unseen SR from 64.1 to 68.3 and SPL from 60.4 to 65.2 over PPO, gains of 4.2 and 4.8 points, respectively.Reward definitions, regularization, output schema, and training budget are controlled in the comparison.Source: Table 16 · p. 19, Appendix A · p. 19

Effect of optimization objective

  • GRPO performs best among the tested optimization objectives, achieving 68.3 SR and 65.2 SPL and outperforming PPO by 4.2 and 4.8 points, respectively.Reward-SFT improves over standard SFT, and policy-based optimization provides further gains; the authors report benefits from sequence-level reward optimization beyond token-level reward-weighted supervision.Source: Table 17 · p. 19, Appendix A · p. 19

APPENDIX B COT DATA GENERATION AND QUALITY CONTROL

The CoT data pipeline separates annotation generation from evaluation data and applies automatic checks followed by manual verification. Filtering yields 134K retained annotations across three reasoning granularities.

Data split integrity

  • CoT annotations use only official training splits of R2R, RxR, and QUARD; validation and test episodes receive no synthetic rationales or action annotations.Evaluation uses original validation splits, human-provided instructions, and official protocols; the teacher receives instructions, observations, and state-action history, but not evaluation targets.Source: Appendix B · pp. 19–20

Dataset composition

  • The final dataset contains 134K annotations: 18K episode-level, 78K step-level, and 38K navigation-level samples.The filtering pipeline reduces 168K raw teacher generations to 134K retained annotations; the granularities cover long-horizon task interpretation, local decisions, and executable task-level actions.Source: Table 18 · p. 20, Appendix B · p. 20

Automatic filtering and manual verification

  • Automatic checks reject malformed structured outputs, missing fields, invalid locomotion values or behavior labels, and irrelevant or unsafe generations; manual review checks rationale-observation and action-reasoning consistency and removes hallucinated or erroneous annotations.Unambiguous formatting or unit-normalization errors may be corrected; manual verification removes approximately 5K additional samples.Source: Appendix B · p. 20
Loading 2609.06251v1…