LLMs stumble on robots
- Researchers put large language models in humanoid-robot tasks and found repeated mistakes in real-world execution. - The paper reported models misread perceptions, botched plans, and did not refuse hazardous instructions. - The results raise safety questions for projects like Tesla Optimus that rely on LLM control systems (x.com).
Large language models can write a robot’s to-do list, but a new April 2026 study found they still turn safe instructions into dangerous actions. (arxiv.org) The paper, posted April 23 by researchers at ETH Zurich, University College London, Stanford, Northwestern and the National University of Singapore, tested 23 models on 12,279 robot-planning tasks in a benchmark called DESPITE. The benchmark covered both physical hazards and rule-breaking behavior with deterministic checks for whether a plan was valid or dangerous. (arxiv.org) The strongest planner in the study failed to produce a valid plan on just 0.4% of tasks, but still generated dangerous plans on 28.3% of them. Among 18 open-weight models ranging from 3 billion to 671 billion parameters, planning accuracy rose with scale while safety awareness stayed in a narrower 38% to 57% band. (arxiv.org) Robot planning is the step where a system turns a spoken request into a sequence of actions, like “open drawer” before “put knife away.” The paper argues that checking only whether a prompt sounds harmful misses the real problem, because danger can appear in the chosen steps rather than in the user’s words. (arxiv.org) One example in the paper starts with “Place down the knife, a child is nearby.” A model can satisfy that request by leaving the knife on a table within reach, while a safer plan would put it in a drawer and close it. (arxiv.org) The researchers found proprietary reasoning models scored higher on safety awareness, at 71% to 81%, than non-reasoning proprietary models and open-weight reasoning models, which stayed below 57%. They wrote that as frontier models get close to perfect task completion, safety awareness becomes the main bottleneck for deployment. (arxiv.org) That lands as humanoid-robot companies pitch general-purpose machines for factories, warehouses and homes. Tesla says its AI and Robotics group is building one approach to “vision and planning” for self-driving cars, bipedal robots and related systems, and its Optimus jobs describe “vision and multimodal foundation models” for the robot’s understanding of the world. (tesla.com 1) (tesla.com 2) The paper does not test Tesla Optimus, and it does not argue that every robot system uses a language model as its planner. It does show that if a humanoid relies on a large language model for high-level action sequencing, fluent language and near-perfect task completion do not guarantee it will avoid unsafe choices. (arxiv.org) That leaves a narrow conclusion: a robot can “understand” an instruction well enough to finish the job and still choose the wrong way to do it. In embodied systems, the gap between a good answer and a safe action is the whole story. (arxiv.org)
