Researchers have found a way to get an off-the-shelf language model to control two robot arms at once, without any task-specific training.
The framework, called BiCICLe, tackles bimanual manipulation by splitting the problem into a leader-follower relationship: one arm's planned action is predicted first, then fed back into the same language model as context for predicting the second arm's move. That sidesteps the core problem with applying in-context learning to two-armed robots: the combined joint action space is too large and too tightly coupled to fit cleanly into a text prompt. Tested on 13 tasks from the TWIN benchmark, BiCICLe hit a 70.5% average success rate, beating the best other training-free method by 6.1 percentage points and outperforming most approaches that were trained specifically for the job. The team also ran it on 3 real-world tasks without retraining for the specific hardware.
That last point matters most. Most capable robot-manipulation systems still need per-task or per-robot training data, which is expensive to collect and transfers poorly between setups. A method that gets a general-purpose language model to coordinate two arms through prompting alone, and still beats specialized, trained systems on most tasks, suggests the gap between generalist reasoning and physical dexterity is narrower than it looks.
Still, treat the real-world numbers with caution: the hardware test covered just 3 tasks, not the full 13-task benchmark, and beating most supervised methods is not the same as beating all of them.