Tesla Chief Admits Optimus Demos Are Remote-Controlled

Ashok Elluswamy, Tesla’s Vice President of AI Software, has offered a candid assessment of the current state of the company’s humanoid robot, Optimus. Speaking on September 20, he clarified that the impressive movements seen in public demonstrations are not the result of independent artificial intelligence. Instead, he described the actions as "essentially remote controlled," a statement that significantly tempers the narrative of immediate autonomy often associated with the technology.

This admission provides a crucial reality check for consumers and investors who may interpret polished demo videos as proof of a finished product. Elluswamy’s explanation highlights a significant gap between the robot’s ability to execute specific, directed tasks and its capacity to navigate unstructured environments on its own. The distinction is vital for understanding the actual technological hurdles remaining in the field of physical robotics.

Remote control versus true autonomy

The term "remote controlled" does not imply that a human is manually operating every joint like a video game character. Rather, a human operator provides high-level instructions, such as picking up an object or walking to a specific location. The robot’s onboard systems then handle the complex, moment-to-moment physics required to maintain balance and coordinate joint movements. This hybrid approach allows the machine to execute instructions while managing its own stability.

However, this setup means Optimus is not currently perceiving its environment and deciding on its own what to do next in a closed loop. It is following a script rather than improvising. This is a meaningful limitation. The hard problems in artificial intelligence lie in the transition from following commands to acting autonomously in unpredictable, real-world settings where the robot must reason about its actions without direct human guidance.

Physical safety presents unique challenges

Elluswamy also drew a sharp contrast between the safety concerns of text-based AI models and those of physical robots. While the industry focuses heavily on issues like hallucinations and alignment in large language models, the stakes are fundamentally different for a machine that interacts with the physical world. In a text-based system, an error results in incorrect information. In a physical robot, an error can result in physical force being applied to a person or object.

The technical challenges include managing the latency between a robot making contact with an unexpected obstacle and reducing its output force. In a home or factory setting, this delay window is critical for preventing injury. The system must process vast amounts of sensory data, potentially over two billion tokens in thirty seconds, and compress it into precise, low-latency physical outputs. This requires response times comparable to or faster than those used in autonomous driving, but with much higher consequences for each decision.

Development stage rather than final product

Elluswamy took over leadership of the Optimus program in June 2025, bringing experience from Tesla’s Full Self-Driving software team. The current architecture relies on a camera-centric approach similar to that used in autonomous vehicles, adapting neural networks to control a body that walks and grasps. While the robot is not yet fully autonomous, this represents a specific development stage rather than a failure. The company’s strategy mirrors its path with self-driving cars, where remote assistance and supervised autonomy preceded broader deployment.

For readers following the progress of humanoid robotics, the key takeaway is that the technology is still in a foundational phase. The ability to balance and execute directed tasks is a significant engineering achievement, but it does not equate to general intelligence. The path to a robot that can safely and independently perform useful work in human environments remains a complex technical challenge, requiring breakthroughs in real-time physics processing and safety assurance.