Interview
Fully autonomous robots are much closer than you think – Sergey Levine
- AI deployments are projected to require hundreds of gigawatts of power with annual marginal capital expenditures reaching $2 to $4 trillion by 2030.
- Robotic foundation models aim to control any robot to perform any task, transitioning from current basic capabilities like laundry folding to fully autonomous, continuous home management over six-month to one-year durations.
- Deployment strategies favor early release of competent robots to initiate a data collection flywheel, with initial useful deployments expected within one to two years and the flywheel starting within single-digit years.
- A median estimate projects fully autonomous home-running robots capable of most blue-collar work will be available within five years, generating productivity gains through expert augmentation rather than wholesale replacement.
- Technological progress relies on a human-in-the-loop approach for bootstrapping, leveraging real-world learning, error correction, and language instructions to expand task scope from specific competencies to general housekeeping.
- Physical robotics offers natural opportunities for improvement through mistake recovery and reflection, contrasting with the supervision signal challenges found in Large Language Model (LLM) human loops.
- Foundation models are expected to achieve compositional generalization, allowing robots to acquire new capabilities like picking up fallen objects without specific training data for those scenarios.
- Current robotics datasets are estimated to be one to two orders of magnitude smaller than multimodal training sets, with the primary challenge being identifying the specific data volume required to sustain a self-driving flywheel.
- Hardware architectures plan to combine pre-trained VLMs with an added "action expert" using flow matching to process sensory information and output continuous actions, while future systems may utilize on-board or off-board inference based on reliability needs.
- Perception capabilities in 2025 are considered a significantly improved starting point compared to 2009, with robotics gaining an advantage through purpose-driven "tunnel vision" compared to the compressed pixel representation of video models.
- Future development requires an industrial-scale building effort focused on foundation models rather than fundamental research, addressing the need to correlate scaling axes with capability improvements to ensure practical utility.
- Trade-offs involving inference speed, context length (ranging from hours to decades), and model size (trillions of parameters) must be resolved, potentially through externalized thinking or multimodal representations that discard irrelevant information.
- Robot arms are expected to drop to hundreds of dollars by 2030 due to economies of scale and AI reducing hardware precision requirements, though current commercial deployment for general-purpose tasks remains very limited.
- Scaling to billions or millions of robots will necessitate solving manufacturing bottlenecks, with significant supply chain concentration in China posing potential risks for US and allied production.
- The robot economy is projected to mature by 2030 to assist with the capital expenditure boom for data centers, solar farms, and chip foundries, potentially enabling construction in remote locations with lower maintenance needs.
- Economic outcomes anticipate a society with full automation and increased wealth, creating a need for redistribution strategies and prioritizing education as the primary lever for workforce flexibility.