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

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

cs.CVSep 5, 2026

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

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

Organizations: School of Computer Science, Peking University, Beijing 100871, China · South China University of Technology, Guangzhou 510006, China · School of Computing, National University of Singapore, Singapore 117417

Abstract

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.

Figures & tables

Fig. 1: Multi-granularity CoT data engine. Given multimodal observations, language instructions, and optional state–action histories, the engine generates episode-, step-, and navigation-level reasoning traces together with executable targets, followed by automatic parsing and semi-automatic quality verification.

Dataset	Nav.	Emb-Ctl.	CoT	Samples
R2R [ 1 ]	✓	✗	✗	50K
RxR [ 20 ]	✓	✗	✗	58K
QUARD [ 39 ]	✗	✓	✗	262K
MobileVLA-CoT-Episode	✗	✓	✓	18K
MobileVLA-CoT-Step	✗	✓	✓	78K
MobileVLA-CoT-Nav	✓	✗	✓	38K

TABLE I: Statistics of source datasets and the synthesized MobileVLA-CoT. Nav.” and Emb-Ctl.” denote navigation and embodied continuous-control supervision, respectively, while CoT” indicates synthesized reasoning annotations. Samples” denotes the number of instances reported by the corresponding datasets.

Fig. 2: Overview of MobileVLA-R1 2.0. Multimodal observations and language instructions are fused for structured reasoning, which conditions a task-level action decoder to predict locomotion $(V_{x},V_{y},\omega)$ and a discrete task-level behavior primitive $\alpha$ for mobile robot control.

Fig. 3: GRPO-based reasoning-to-action optimization. For each multimodal input, the policy samples multiple structured outputs whose induced actions are evaluated by movement, behavior, and format rewards. Group-relative advantages, together with KL regularization to a frozen reference policy, are then used for offline policy optimization.

Method	Observation				R2R-CE Val-Unseen				RxR-CE Val-Unseen
Method	S.RGB	Pano.	Depth	Odom.	NE $\downarrow$	OS $\uparrow$	SR $\uparrow$	SPL $\uparrow$	NE $\downarrow$	SR $\uparrow$	SPL $\uparrow$	nDTW $\uparrow$
CMA ∗ [ 49 ]	✗	✓	✓	✓	6.20	52.0	41.0	36.0	8.76	26.5	22.1	47.0
Sim2Sim ∗ [ 50 ]	✗	✓	✓	✓	6.07	52.0	43.0	36.0	-	-	-	-
GridMM ∗ [ 51 ]	✗	✓	✓	✓	5.11	61.0	49.0	41.0	-	-	-	-
Ego 2 -Map ∗ [ 52 ]	✗	✓	✓	✓	5.54	56.0	47.0	41.0	-	-	-	-
DreamWalker ∗ [ 53 ]	✗	✓	✓	✓	5.53	59.0	49.0	44.0	-	-	-	-

TABLE II: Comparison on VLN-CE Val-Unseen [ 48 , 20 ] . We report standard metrics on the R2R-CE and RxR-CE val-unseen splits. “S.RGB”, “Pano.”, and “Odo.” denote single-view RGB, panoramic observations, and odometry, respectively. ∗ indicates the use of a simulator-pretrained waypoint predictor [ 48 ] .

Fig. 4 : Qualitative results on R2R-CE. MobileVLA-R1 2.0 executes navigation instructions in the simulator.

Fig. 5 : Qualitative results on RxR-CE. MobileVLA-R1 2.0 follows multilingual instructions with complex spatial semantics, maintaining coherent reasoning and stable motion across unseen environments.

Fig. 6: Real-world hardware and deployment pipeline on Unitree Go2. (a) Go2 platform equipped with an Intel RealSense D435i RGB-D camera, an L2 LiDAR, and a Jetson Orin Nano. (b) Hybrid Deployment pipeline. Sensing and multimodal preprocessing are performed onboard, while the 8B VLA backbone and reasoning-conditioned action decoder are executed on a remote H20 GPU. The resulting task-level commands $(V_{x},V_{y},\omega,\alpha)$ are transmitted back to the robot.

Fig. 7: Real-world hardware and deployment pipeline on Unitree G1. (a) G1 humanoid platform equipped with an Intel RealSense RGB-D camera, an Inspire dexterous hand, and a Jetson Orin NX computation module. (b) Hybrid Deployment pipeline. RGB-D observations and language instructions are processed by MobileVLA-R1 2.0, where the reasoning-conditioned action decoder maps multimodal representations to task-level commands.

Fig. 8: Indoor qualitative results on Unitree Go2. Representative closed-loop rollouts show target-directed navigation, turning, and obstacle avoidance in indoor environments.

Fig. 9: Outdoor qualitative results on Unitree Go2. Representative closed-loop rollouts show target-directed navigation, turning, and obstacle-aware execution in outdoor environments.

Fig. 10: Real-world qualitative results on Unitree G1. Representative mobile-manipulation rollouts in the Cluttered , Shelf/Cabinet , and Tabletop scenarios, showing task progression from navigation and alignment to grasping and object interaction.

Method	Easy	Medium	Hard				Average
Method	Distinguish	Go-to	Go-avoid	Go-through	Crawl	Unload	Average
CLIP [ 67 ]	0.44	0.43	0.45	0.19	0.00	0.00	0.25
VC-1 [ 68 ]	0.46	0.43	0.45	0.31	0.00	0.00	0.28
QUART [ 39 ]	0.66	0.60	0.53	0.41	0.32	0.12	0.44
MoRE [ 17 ]	0.82	0.80	0.59	0.57	0.49	0.33	0.60
MobileVLA-R1 [ 19 ]	0.92	0.89	0.71	0.65	0.58	0.44	0.70

TABLE III: Overall performance on QUARD [ 39 ] . We report success rates on six tasks grouped by difficulty; Average is the mean over all tasks, with 25 episodes per task.

Method	Workspace				Corridor				Outdoor
	Simple		Complex		Simple		Complex		Simple		Complex
	NE $\downarrow$	SR $\uparrow$	NE $\downarrow$	SR $\uparrow$	NE $\downarrow$	SR $\uparrow$	NE $\downarrow$	SR $\uparrow$	NE	SR $\uparrow$	NE	SR $\uparrow$
GPT-4o [ 72 ]	2.01	0.67	2.38	0.33	1.49	0.53	3.00	0.00	–	0.67	–	0.50
NaVILA [ 47 ]	1.29	0.83	1.76	0.80	1.15	0.89	1.76	0.67	–	0.91	–	0.83
MobileVLA-R1 [ 19 ]	1.03	0.93	1.23	0.91	0.96	1.00	1.23	0.86	–	1.00	–	0.96
MobileVLA-R1 2.0 (Ours)	0.96	0.95	1.12	0.94	0.90	1.00	1.11	0.91	–	1.00	–	0.98

TABLE IV : Closed-loop real-world evaluation on Unitree Go2. We report navigation error (NE) and success rate (SR) under Simple and Complex instructions in three representative environments. Outdoor results use SR only because reliable global localization is unavailable.

Setting	Tasks	Trials	Episodes	Latency (ms)	Failure breakdown
Workspace-Simple	6	5	30	$210\pm 14$	Target 1 / Percep. 1
Workspace-Complex	6	5	30	$218\pm 16$	Target 2 / Percep. 1
Corridor-Simple	5	5	25	$205\pm 13$	–
Corridor-Complex	5	5	25	$216\pm 17$	Turn 2 / Pass. 2
Outdoor-Simple	5	5	25	$232\pm 21$	–
Outdoor-Complex	5	5	25	$245\pm 24$	Drift 1

TABLE V: Deployment statistics and failure analysis on Unitree Go2. Latency is reported as the mean

\pm

standard deviation of the hybrid end-to-end latency measured from observation acquisition to task-level command handoff, and failures are counted at the episode level. “Target,” “Percep.,” “Turn,” “Pass.,” and “Drift” denote target-grounding, perception/obstacle, excessive-turning, narrow-passage, and localization-drift failures, respectively.

Method	Tabletop			Shelf/Cabinet			Cluttered			Full Avg.
Method	Nav.	Manip.	Full	Nav.	Manip.	Full	Nav.	Manip.	Full	Full Avg.
NaVILA [ 47 ]	75.0	63.3	47.5	67.5	59.3	40.0	57.5	43.5	25.0	37.5
MobileVLA-R1 [ 19 ]	80.0	71.9	57.5	72.5	65.5	47.5	65.0	53.8	35.0	46.7
MobileVLA-R1 2.0 (Ours)	85.0	79.4	67.5	77.5	74.2	57.5	70.0	64.3	45.0	56.7

TABLE VI: Real-world mobile manipulation on Unitree G1. Navigation (Nav.), manipulation (Manip.), and full-task (Full) success rates are reported in percentage. Manip. denotes the manipulation success rate conditioned on successful navigation, whereas Full requires both navigation and manipulation to succeed within the same episode. All methods use the same sensing and low-level control interfaces, with no G1-specific policy fine-tuning.

Scenario	Tasks	Trials/Task	Episodes	Successful	Failure Breakdown
Tabletop	4	10	40	27	Ground. 2 / Nav. 2 / Grasp 5 / Manip. 3 / Ctrl. 1
Shelf/Cabinet	4	10	40	23	Ground. 2 / Nav. 3 / Grasp 5 / Manip. 6 / Ctrl. 1
Cluttered	4	10	40	18	Ground. 4 / Nav. 5 / Grasp 6 / Manip. 5 / Ctrl. 2
Total	12	–	120	68	Ground. 8 / Nav. 10 / Grasp 16 / Manip. 14 / Ctrl. 4

TABLE VII: Evaluation scale and failure statistics on Unitree G1. Failures are counted at the episode level and categorized according to their dominant observed cause. “Ground.,” “Nav.,” “Grasp,” “Manip.,” and “Ctrl.” denote target-grounding, navigation/positioning, grasping, manipulation-execution, and low-level control failures, respectively.

Action Interface	Obs.	Reason.	R2R-CE		RxR-CE
Action Interface	Obs.	Reason.	SR $\uparrow$	SPL $\uparrow$	SR $\uparrow$	SPL $\uparrow$
Deterministic Parsing	✗	✗	68.3	65.2	71.5	66.8
Learnable Decoder	✓	✗	68.8	65.7	72.0	67.3
Learnable Decoder	✗	✓	69.3	66.3	72.6	67.9
Learnable Decoder	✓	✓	69.8	66.9	73.1	68.5

TABLE VIII: Ablation of action decoding and conditioning. We compare deterministic action parsing with learnable action decoders under different conditioning inputs. “Obs.” and “Reason.” denote observation and intermediate reasoning representations, respectively.

Action Decoder	GRPO	R2R-CE		RxR-CE
Action Decoder	GRPO	SR $\uparrow$	SPL $\uparrow$	SR $\uparrow$	SPL $\uparrow$
✗	✗	58.0	53.2	61.4	56.7
✓	✗	63.1	58.7	66.0	61.4
✗	✓	68.3	65.2	71.5	66.8
✓	✓	69.8	66.9	73.1	68.5

TABLE IX: Ablation of the action decoder and GRPO optimization. We evaluate the individual and joint effects of the two components. Without the action decoder, task-level actions are obtained by deterministic parsing; without GRPO, the model is trained with supervised learning only.

Reasoning Granularity	SR $\uparrow$	SPL $\uparrow$
No-CoT	64.0	59.6
Episode-level	65.5	61.3
Step-level	66.0	61.7
Navigation-level	66.7	62.6
Multi-granularity	68.3	65.2

TABLE X: Ablation of reasoning supervision granularity on R2R-CE Val-Unseen. We vary the granularity of reasoning supervision while keeping the teacher model, action targets, output format, and training budget fixed. No-CoT removes rationale supervision while preserving identical action supervision.

$R_{\rm mov}$	$R_{\rm beh}$	$R_{\rm fmt}$	SR $\uparrow$	SPL $\uparrow$
✗	✗	✗	58.0	53.2
✓	✗	✗	60.7	55.5
✗	✓	✗	61.9	56.8
✗	✗	✓	59.6	54.7
✓	✓	✗	64.5	60.2
✓	✗	✓	63.4	59.1

TABLE XI: Ablation of GRPO reward components on R2R-CE Val-Unseen. Results are obtained with deterministic action parsing. We evaluate the movement, behavior, and format rewards; the configuration without reward optimization denotes the supervised pre-GRPO baseline.

Action Decoder	SR $\uparrow$	SPL $\uparrow$
Mean Pooling + MLP	68.7	65.6
Single-Query Attention	69.1	66.1
Shared-Query Dual-Head	69.4	66.5
Dual-Query Dual-Head	69.8	66.9

TABLE XII: Ablation of action decoder architectures on R2R-CE Val-Unseen. All variants use identical observation and reasoning representations and the same training protocol, differing only in the decoder architecture.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	R2R-CE		RxR-CE
Setting	SR $\uparrow$	SPL $\uparrow$	SR $\uparrow$	SPL $\uparrow$
Text + RGB	62.5	59.0	66.0	61.5
+ Depth	66.0	62.0	69.0	64.0
+ Point Cloud	67.2	63.5	70.2	65.2
MobileVLA-R1 (Full)	68.3	65.2	71.5	66.8

Appendix

TABLE XIII: Incremental ablation of multimodal perception on VLN-CE. The first three rows progressively add geometric modalities under the same training configuration. The final row reports the complete MobileVLA-R1 configuration for reference.

$\lambda_{\rm mov}$	SR $\uparrow$	SPL $\uparrow$	Adv. Contrib. (%)
$\lambda_{\rm mov}$	SR $\uparrow$	SPL $\uparrow$	$R_{\rm mov}$	$R_{\rm beh}$	$R_{\rm fmt}$
0.00	65.2	61.0	0	62	38
0.10	66.4	62.4	8	58	34
0.25	67.1	63.5	16	54	30
0.50	67.8	64.4	24	50	26
1.00	68.3	65.2	32	46	22

Appendix

TABLE XIV: Sensitivity to the movement reward weight on R2R-CE Val-Unseen. Results are obtained with deterministic action parsing. We vary

\lambda_{\rm mov}

while keeping

\lambda_{\rm beh}

and

\lambda_{\rm fmt}

fixed. “Adv. Contrib.” denotes the normalized contribution of each reward term to the reward-derived advantage, with all contributions summing to 100%.

Rationale Source	SR $\uparrow$	SPL $\uparrow$
No-CoT	64.0	59.6
Template-CoT	65.1	60.8
LLaMA-3-8B CoT	66.2	62.0
Gemini-CoT	68.3	65.2

Appendix

TABLE XV: Ablation of rationale sources on R2R-CE Val-Unseen. All variants use identical action targets, backbone, structured output format, and training budget; only the source of supervised rationales is varied.

Optimization Method	SR $\uparrow$	SPL $\uparrow$
PPO	64.1	60.4
GRPO	68.3	65.2

Appendix

TABLE XVI: Comparison of PPO and GRPO on R2R-CE Val-Unseen. Both methods use the same reward definitions, structured output schema, KL regularization, and training budget.

Method	Sampling	Reward	Optimization Objective	SR $\uparrow$	SPL $\uparrow$
SFT	✗	✗	Token-level CE	58.0	53.2
Reward-SFT	✗	✓	Reward-weighted CE	62.4	58.3
PPO	✓	✓	Policy Gradient	64.1	60.4
GRPO	✓	✓	Group-relative Objective	68.3	65.2

Appendix

TABLE XVII: Comparison of optimization objectives on R2R-CE Val-Unseen. All variants use the same training data and action targets. Reward-based variants use identical reward definitions and deterministic action parsing.

Filtering Stage	Input	Retained	Ret.	Primary Removal Reason
Raw generation	168K	168K	100.0%	–
Format validation	168K	158K	94.0%	malformed tags / missing fields
Action validation	158K	146K	92.4%	invalid commands / out-of-range values
Safety and relevance filtering	146K	139K	95.2%	unsafe or instruction-irrelevant outputs
Manual verification	139K	134K	96.4%	hallucination / action mismatch / visual inconsistency
Final	134K	134K	100.0%	18K episode / 78K step / 38K navigation

Appendix

TABLE XVIII: Filtering statistics of MobileVLA-CoT. “Ret.” denotes the fraction of samples retained from the preceding stage.

Language-Conditioned Robot Control

Language-conditioned robot control connects natural-language instructions to grounded, safe actions across tasks and robot embodiments.

See all 74 papers

ReFineVLA: Multimodal Reasoning-Aware Generalist Robotic Policies via Teacher-Guided Fine-Tuning
Apr 20, 2026Tuan Van Vo, Tan Q. Nguyen, Khang Nguyen +5Robotic Manipulation VLM Reasoning
VinRobotics, Hanoi. Vietnam · Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), UAE. · University of Texas at Arlington, Texas, USA +4
H-VLA: Hierarchical Vision-Language-Action Model with Key-Action Reasoning and Motion Planning in a Unified Action Space
Sep 19, 2026Xiongfeng Peng, Lu Xu, Yandong Wang +10Vision-Language-Action Models Language-Conditioned Robot Manipulation
Advanced Research Lab, Samsung R&D Institute China-Beijing (SRCB), China · Samsung AI Center, DS Division, South Korea
Guide, Think, Act: Interactive Embodied Reasoning in Vision-Language-Action Models
May 13, 2026Yiran Ling, Qing Lian, Jinghang Li +6Robotic Control Robot Policy Learning
Faculty of Computing, Harbin Institute of Technology · International Digital Economy Academy (IDEA) · National Key Laboratory of Smart Farm Technologies and Systems +4

Apr 20, 2026J/K move · Enter open · S save

ReFineVLA: Multimodal Reasoning-Aware Generalist Robotic Policies via Teacher-Guided Fine-Tuning

Tuan Van Vo, Tan Q. Nguyen, Khang Nguyen, Nhat Xuan Tran+4

VinRobotics, Hanoi. Vietnam · Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), UAE. · University of Texas at Arlington, Texas, USA+4

Robotic Manipulation VLM Reasoning Vision-Language-Action Models Reasoning for VLA Models

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models have gained much attention from the research community thanks to their strength in translating multimodal observations with linguistic instructions into desired robotic actions. Despite their advancements, VLAs often overlook explicit reasoning and learn the functional input-action mappings, omitting crucial logical steps, which are especially pronounced in interpretability and generalization for complex, long-horizon manipulation tasks. In this work, we propose ReFineVLA, a multimodal reasoning-aware framework that fine-tunes VLAs with teacher-guided reasons. We first augment robotic datasets with reasoning rationales generated by an expert teacher model, guiding VLA models to learn to reason about their actions. Then, we fine-tune pre-trained VLAs with the reasoning-enriched datasets with ReFineVLA, while maintaining the underlying generalization abilities and boosting reasoning capabilities. We also conduct attention map visualization to analyze the alignment among visual observation, linguistic prompts, and to-be-executed actions of ReFineVLA, reflecting the model is ability to focus on relevant tasks and actions. Through this additional step, we explore that ReFineVLA-trained models exhibit a meaningful agreement between vision-language and action domains, highlighting the enhanced multimodal understanding and generalization. Evaluated across a suite of simulated manipulation benchmarks on SimplerEnv with both WidowX and Google Robot tasks, ReFineVLA achieves state-of-the-art performance, in success rate over the second-best method on the both the WidowX benchmark and Google Robot Tasks.

Abstract

Language-Conditioned Robot Control

Explore similar work

ReFineVLA: Multimodal Reasoning-Aware Generalist Robotic Policies via Teacher-Guided Fine-Tuning

H-VLA: Hierarchical Vision-Language-Action Model with Key-Action Reasoning and Motion Planning in a Unified Action Space

Guide, Think, Act: Interactive Embodied Reasoning in Vision-Language-Action Models