MobileVLA-R1 2.0: RL-Enhanced Reasoning for Mobile Robot Control
The paper claims better robot instruction-following by tying explicit reasoning to executable control.
MobileVLA-R1 2.0 uses supervised chain-of-thought alignment plus reinforcement learning to improve consistency between embodied reasoning and robot actions. Its action decoder maps multimodal reasoning into task-level targets, leaving robot-specific controllers to handle actuation. The authors report gains over MobileVLA-R1, including +1.6 success rate points on VLN-CE and +10.0 points on real-world G1 mobile manipulation tasks. HF Daily Papers' note
MobileVLA-R1 2.0 uses supervised chain-of-thought alignment plus reinforcement learning to improve consistency between embodied reasoning and robot actions. Its action decoder maps multimodal reasoning into task-level targets, leaving robot-specific controllers to handle actuation. The authors report gains over MobileVLA-R1, including +1.6 success rate points on VLN-CE and +10.0 points on real-world G1 mobile manipulation tasks. HF Daily Papers' note
score 5