Humanoid Robots are a Vehicle for AGI; Teleoperation Creates Partial Viability for Now

Humanoid robot uptake continues to disappoint and the hype continues to build with a new vendor joining the race practically every fortnight. It is hard to deny that in our arrogance, humans have designed the “ultimate robot” in our own image. It is also hard to deny that the intelligence humanoid robots ought to be endowed with to take on the roles of human workers—logic, adaptation, spatial awareness—are defining attributes of Artificial General Intelligence (AGI)—technology that does not yet (and may never) exist.

Current Artificial Intelligence (AI) is inept at dealing with the real world. Multi-modal foundation models, including NVIDIA’s GR00T and The Toyota Research Institute’s Large Behavioral Model, have demonstrated the ability to program robots (not just humanoids) for novel tasks like tying shoelaces or picking heterogeneous objects from tabletops via brute-force (simulation based) reinforcement learning—with impressive, if clunky, results. But this is not a practically viable strategy for achieving machine autonomy at scale. Reinforcement learning (using simulation, synthetic data, and shed loads of compute) will fall short because the approach is tantamount to simulating every permutation of every scenario millions of times until convergence is achieved and a robot knows “what to do”—manually programming the robot to perform a dexterous task would be a lot faster and cheaper. The reinforcement learning approach alone is doomed to fail. There are simply too many variables in the real world for robots to not be thrown off task when confronted with a slight variant of the trained scenario—look no further than the failure rate of autonomous vehicles. Reality presents too many edge cases for simulation to offer adequate predictability or repeatability. More resources need to be dedicated to edge-compute and inference.

Humanoid robotics has come a long way since the Defense Advanced Research Projects Agency (DARPA)-funded Atlas robot was unveiled in 2013. Key innovations include stability, hand dexterity, and low-latency teleoperation.

A focus on hand dexterity is important for unlocking new tasks and associated Return on Investment (ROI)—without the ability to perform complex manual tasks, humanoid robots are simply over-engineered Autonomous Mobile Robots (AMRs), good for material handling, but little else. Many organizations focus on dexterous—hand-like—manipulators for humanoids. For example, Sanctuary AI is a market leader for humanoid hand dexterity with over 200 patents filed or pending. Again, the use cases for hands are currently stymied by regulation, safety, and repeatability—the key metric that directly translates to value.

All humanoids utilize teleoperation. There is always a human operator in the background in case things go wrong (generally, future funding and deployment opportunity depend on this). The uncannily human behavior of Optimus at Tesla’s 2024 AI Day led skeptics to compliment the low latency of the device’s teleoperation. Teleoperation for robotics in high-risk deployments is not new. InOrbit and Brain Corp offer teleoperation as part of their offering for AMRs to support deployments in public spaces where stakeholders cannot allow robots to get lost or drift too far from their paths. Agility Robotics’ humanoids all have human oversight in the form of teleoperation; Apptronik applies teleoperation to its Apollo robot for teaching by demonstration. The hardware is impressive and, before long, it’ll be possible to be in two places at once. Teleoperation advances create new commercial opportunities for humanoids, including the telepresence of individuals for customer service; inspection and repair in dangerous or remote locations; and assistive robots within the healthcare sector.

However, more fundamental software technologies have not advanced in step with humanoid hardware. Simultaneous Localization and Mapping (SLAM), a software methodology for enabling robots to navigate autonomously, remains an unsolved engineering problem. Consequently, humanoids will still get lost, autonomous cars will veer into traffic under certain bad conditions, and autonomous tractors will drift from a plow line in an empty field. Emerging technologies, such as quantum magnetometers and photonic Inertial Measurement Units (IMUs), will improve SLAM, but they won’t completely solve it—noise and drift are fundamental features of sensors.

Despite the push from industry giants like NVIDIA, increasing the amount of computational resources dedicated to training robots will not create markedly more capable solutions—robots will still be dependent on less than perfect sensors, flawed inference capabilities due to the unpredictability of reality, and a lack of higher comprehension of the task given to them.

That is not to say that this current stage of humanoid development is without value. Stakeholders may yet prove sufficient value for humanoids within manufacturing verticals. Automotive, in particular, offers well controlled and monitored environments with sufficient complexity and uniformity for humanoids to demonstrate value, while minimizing repeatability concerns. Several humanoid manufacturers recognize the value of humanoids within automotive manufacturing and decision makers have been receptive. For example, Sanctuary AI has a partnership with Magna International for manufacturing car parts; Optimus can be used to manufacture Teslas; Boston Dynamics will use Atlas to assemble Hyundai cars; Apptronik and Fourier are tackling Mercedes and BMW, respectively; in China, the Electric Vehicle (EV) assembler Nio is running trials with UBTECH’s Walker S and Leju Robotics’ Kuavo. Perhaps these solutions, with the help of teleoperation, will prove viable at scale for automotive manufacturing; it will be longer before we see humanoids in retail stores among humans.

Alternative approaches to machine intelligence—or even computing in general—may offer a solution. ABI Insight “Biomimicry Outperforms Generative AI for Robotics” outlines the commercial maturity of innovators in this space. Regardless of shortcomings, the demand for humanoid robotics will not abate: the key driving issues of labor shortages and care for aging populations will only increase as the century unfolds.