BEIJING, Sept 2 – Turing Award winner Richard Sutton and Chinese tactile-sensor maker Tashan Technology opened a robot “kindergarten” in Beijing on Tuesday with a wager: the next advance in robotics will come not only from more human demonstrations or simulation, but from machines learning through their own mistakes after deployment. The roughly 10 robots shown to media, however, remained under remote control, China Economic Net reported.
That distinction goes to a central problem in embodied AI. Much robot training turns demonstrations, teleoperation or simulated experience into a policy that is then installed on a machine. Sutton’s project is trying to make learning continue in the physical world, where joints loosen, objects deform and failures can damage hardware.
That makes durability, rather than another algorithm alone, part of the bet. “If it can’t make a mistake, then it can’t learn,” Kris De Asis, a senior research fellow at Openmind, told reporters, identifying robots that break after repeated falls as a near-term bottleneck.
Tashan’s tactile sensors and electronic skin are intended to give the machines contact data and protective reflexes. The two-floor site at Beijing’s Shougang Park has areas for testing, learning and interaction and is recruiting robot makers, sensor companies, algorithm teams and universities, CEN said.
The kindergarten label is more distinctive than the underlying research direction. The Technical University of Munich uses the same name for a project centred on learning from human observation, while a Google Research waste-sorting programme logged 9,527 robot-hours across 23 machines. What could set Beijing apart is sustained learning on deployed robots with limited human intervention, not physical reinforcement learning by itself.
The opening did not show that. De Asis said a small spider-like robot had learned to move forward on physical hardware in “maybe 40 minutes”, but called it a very specific case. Project materials reviewed by SouthwardTech provide no operating schedule, robot-hours, failure or reset counts, or evidence that learning transfers across tasks or machines.
For now, the kindergarten shows where its backers believe the industry’s bottleneck lies. It does not yet show they have cleared it.
Related: Unitree Robotics has priced a STAR Market IPO that raises about 1.7bn yuan more than its prospectus allocates to projects.
