DeepMind and Anthropic Explore Embodied Reasoning: General Models Still Can't Drive Robots Alone

RanjayKrishna · x · 2026-08-01

Recent developments from Google DeepMind and Anthropic have sparked discussions on whether reasoning models can directly decide a robot's whole-body movements. DeepMind's Gemini Robotics 2 layers an embodied reasoning model atop a VLA for whole-body control, while Anthropic has tested general-purpose reasoning models at various control levels.

The consensus is clear: general-purpose reasoning models still cannot control humanoids independently and require a pretrained policy underneath. However, researchers believe that strong multimodal and reasoning capabilities could eventually make this a reality.

To explore this, researchers introduced HumanCLAW. The work decouples 'Action Intelligence' from motor control to finally measure it, revealing that today’s best models are surprisingly bad at it.

Original post →

More from Embodied

Embodied channel →