Jitendra Malik is asking artificial intelligence to learn from the world before it tries to describe it. His 2026 research and teaching agenda at the University of California, Berkeley has emphasised sensorimotor learning: systems that connect perception, action and physical consequence. The idea challenges an industry that has concentrated enormous capital on models trained primarily through language.
Text-trained models can reason over instructions, code and recorded knowledge, but a robot must handle friction, balance, occlusion, uncertainty and the consequences of a bad movement. It needs to know not only what a cup is, but how firmly to hold one, what happens when it is full and when to stop. Those capabilities emerge through interaction.
Malik, the Arthur J. Chick Professor at Berkeley and a foundational figure in computer vision, leads through research, education and the people he has trained. His focus on embodied intelligence can influence where laboratories and companies invest. The central business question is whether robotics can build a data economy comparable to the web-scale resources that powered text-trained models.
Physical experience is expensive
Text and images can be copied and processed at enormous scale. Robot data requires hardware, space, maintenance and supervision. A single grasp can take seconds, and failed actions may damage an object or machine. Different robots produce different sensor and control formats, making datasets difficult to combine.
This scarcity creates a concentration risk. Companies with large fleets can collect more interactions and improve faster, while universities and startups struggle to match them. The market may consolidate around organisations that own both hardware and data before the science is mature.
Malik’s academic position makes shared resources important. Public datasets, standard interfaces and reproducible tasks can keep the field open. They will not replace proprietary fleet data, but they can provide a common foundation for comparison and education.
Data quality matters more than raw hours. Repeated easy motions add little if a system never encounters clutter, failure or recovery. Datasets should include interventions, uncertainty and varied environments. The moments when a human takes over can be especially informative.
Simulation is necessary but not sovereign
Simulation can generate millions of interactions without risking hardware. Researchers can vary lighting, object properties and layouts, train policies and test rare events. Advances in graphics and learned world models make simulated environments increasingly useful.
The gap between simulation and reality remains. Contact dynamics, deformable objects and sensor noise are hard to reproduce. A policy that performs perfectly in a clean virtual kitchen may fail when a bag bends or a camera lens is dirty. Simulation can accelerate learning, but real-world validation determines safety.
Malik’s vision background is relevant because perception systems have long confronted domain shift. Methods that learn invariant structure, combine synthetic and real data and adapt with limited examples can reduce the gap. Robotics needs equivalent discipline for action.
Companies should report how much real-world data supports a claimed capability and where the system was tested. A visually impressive demonstration in a prepared environment is not evidence of general operation. Benchmarks should include novel objects, disturbances and failure recovery.
Action changes the meaning of intelligence
A text-trained model can produce a wrong answer without directly moving equipment. A robot’s error can injure a person, damage inventory or stop a production line. Intelligence must therefore include knowing when not to act and when to request help.
Sensorimotor learning can make uncertainty operational. A system can estimate whether an object is within its experience and choose a lower-risk motion. It can gather information by changing viewpoint or touching gently. These behaviours are closer to scientific experimentation than to next-token prediction.
Evaluation should reward calibrated hesitation and safe recovery, not only task completion. A robot that succeeds 95 per cent of the time but causes severe failures may be less useful than one that completes fewer tasks and escalates appropriately. Commercial buyers need distributions, not highlight reels.
Malik can help define research questions around intervention, exploration and failure. Those questions influence how students and companies design systems. The field needs theory and benchmarks for safe action before autonomy expands into less controlled environments.
Foundation models need bodies and interfaces
Robotics companies are developing models intended to operate across tasks and hardware. A common model can interpret images and language, plan an action and generate controls. The promise is that learning transfers rather than requiring a separate programme for every task.
Transfer is difficult because bodies differ. A two-finger gripper, dexterous hand and mobile manipulator have different capabilities. Sensor placement and control frequency matter. Models need representations that separate the goal and world from the specific machine.
Standard action interfaces could improve portability, much as software standards expanded computing ecosystems. Hardware makers may resist if proprietary control creates differentiation. Buyers, however, benefit when models and data can move among platforms.
Universities can test architectures across varied robots and publish failure modes. Malik’s Berkeley course on robots that learn can train engineers to think across vision, control and machine learning rather than treating each as an isolated specialty.
The labour model should be designed early
Robot learning often depends on people who demonstrate tasks, label data, reset environments and take control during errors. Their work can be hidden behind claims of autonomy. Companies should measure and disclose the human time required per operating hour.
Remote operation may remain part of commercial systems for years. It can make deployments useful while models improve, but economics depend on how many robots one person can supervise and how often intervention occurs. Safety and working conditions for operators matter.
Workers in factories, warehouses and care settings possess practical knowledge that researchers need. Involving them in task design can reveal edge cases and create better roles. Deployment plans should include training and transition rather than treating labour displacement as an external consequence.
Malik’s sensorimotor emphasis supports this view because human demonstration is not merely labelled output; it is embodied expertise. Systems should learn from that knowledge while organisations recognise and compensate it.
Asia is both a market and a research environment
Asia contains advanced manufacturing, logistics networks, ageing societies and rapidly growing service economies. Japan, South Korea, China and Singapore have strong robotics ecosystems, while India and Southeast Asia offer different labour economics and operating conditions. A system proven in one market may not transfer automatically.
Factories provide structured environments and repeatable tasks, making them early adopters. Service and home settings are more variable. Languages, building layouts, objects and social expectations change. Data collection needs regional diversity if embodied models are to generalise.
Partnerships between universities, manufacturers and operators can create test sites with independent evaluation. Data-sharing agreements should protect commercial information and worker privacy while allowing useful findings to circulate.
Malik’s international influence and Berkeley network can connect these communities. The goal should not be one universal dataset controlled by a single company, but interoperable resources that reflect varied bodies and environments.
Robotics needs a credible scaling law
Language-model investment was supported by a visible relationship among data, compute, model size and benchmark performance. Robotics lacks an equally reliable scaling law. More interactions can help, but hardware diversity and physical noise make returns uncertain.
Researchers should test how performance changes with demonstrations, fleet size, task diversity and simulation. Companies need to know whether doubling data improves generalisation or merely repeats known situations. Efficient learning may be more valuable than the largest robot model.
Compute economics also differ. Training can occur centrally, but inference must often run with low latency and limited power. Edge hardware, connectivity and safety systems add cost. A viable product requires the whole system to deliver value above labour and capital expense.
Malik’s insistence on grounding can keep the field focused on real capability. A model that discusses physics well is not a robot that can adapt safely. Investment should follow evidence from sustained operation.
Academic leadership can keep evaluation honest
Universities have incentives and conflicts of their own, but they can publish methods, compare systems and study questions that do not fit a product cycle. Malik’s election as a Fellow of the Royal Society in 2026 recognises a career that connected fundamental vision research with broad application. That standing gives him a platform to demand stronger evidence.
Benchmarks should be resistant to overfitting and updated as systems improve. Test environments can be held out, with independent operators running trials. Safety events and human interventions should be part of the score.
Funding agencies can reinforce that discipline by supporting shared test facilities and long-duration studies, not only new model architectures. Reliable robotics requires maintenance, systems engineering and repeated measurement. Those investments may produce fewer spectacular announcements, but they create evidence that companies and public institutions can use.
Jitendra Malik’s sensorimotor agenda is a reminder that intelligence is not only an answer. It is an action in a world that pushes back. Robotics will scale when companies and researchers can turn expensive physical experience into transferable learning without hiding failure or human support. Building that data economy is less glamorous than a humanoid demonstration, but it will determine which embodied systems become reliable infrastructure.
