On Tuesday 1 September 2026, a robot kindergarten opened in Beijing’s Shijingshan district. Global Times reporter Zhang Yiyi filed the English account the same evening. The facility is jointly established by Chinese tactile-sensing company Tashan Technology and a team led by Turing Award winner and reinforcement-learning pioneer Richard Sutton. Asia Business Daily’s 2 September follow-ups add that CEO Ma Yang said the site opened 37 days after the plan was announced, with support from Beijing city and Shougang Group. Sutton’s own homepage at incompleteideas.net confirms he is working with Tashan Technologies on a robot kindergarten in Beijing via Openmind. That is an opening ceremony and a research thesis, not a published joint map and not a continuous-duty report.
The thesis is tactile sensing plus continuous learning. Robots are meant to learn through touch, trial and error, and their own experience, not only from human demonstrations. Global Times observed machines repeatedly trying movements, taking physical feedback, and adjusting. A wall collision while learning to walk becomes training data rather than a scrubbed failure. Ma Yang framed the starting point as tactile perception in a controlled environment. Sutton told the opening that robots need a safe place to learn their bodies and the physical world through experience, and that this needs new algorithms plus robots and environments designed for learning. Failures are part of the process on that framing. FIRGELLI will treat that as a research charter until someone prints which platforms are inside, what their DoF tables look like, and how many hours a joint can take a hard stop before thermal derate.
What the kindergarten claims. What it does not publish.
Nowhere in the Global Times piece, the Asia Business Daily notes, or Sutton’s homepage note does FIRGELLI find a humanoid DoF count, rated payload at reach, battery watt-hours, continuous hip or knee torque, pack temperature after a thirty-minute stumble session, or a named platform SKU for the bodies on the floor. That absence matters. A kindergarten for continuous learning is only as useful as the hardware that survives the mistakes. Kris De Asis of Openmind Global Research, quoted by Global Times, put the bottleneck in plain language: if a robot cannot make a mistake, it cannot learn. He also noted that many robots still train via demos, teleoperation, or simulation and then ship with behavior largely fixed. Continuous learning aims to keep adapting after deployment. It is slower than imitation. Near-term demos favor teleop and mocap. Long-term autonomy bets on experience. Those are different calendars.
Asia Business Daily names Chris de Assis / Hou Guangdong against seven major joint research tasks. FIRGELLI matches the person to Kris De Asis as Global Times spelled it, and treats the seven-task list as a research agenda rather than a published acceptance test. OpenMind’s earlier curriculum work with Robostore last November is background only. It does not prove Shijingshan has a finished continuous-learning stack on a full-size biped. Treat the seven tasks as a work plan until someone publishes method, metric, and platform.
Spider demo is not a humanoid duty cycle.
Global Times carries a narrow demo from Kris De Asis: a small spider-like robot at the facility learned to move forward in about forty minutes with no prior knowledge. Read that as a low-DoF locomotion toy under a tightly defined reward, not as evidence that a balance-critical humanoid can keep learning after a warehouse shift. Kris himself flagged the gap. Complex humanoids involve balance, energy use, temperature control, and self-defense. A forty-minute crawl on a light multi-leg frame does not transfer to hip and knee modules under stop-and-turn, or to a wrist that has to hold force after a collision. FIRGELLI’s bar stays boring. Continuous torque under repeated falls. Gripper force after a bump. Pack temperature when the learning loop keeps commanding retries. Which joints are rated for impact loads. None of those rows are on today’s public sheet.
Wang Peng of the Beijing Academy of Social Sciences called the work early-stage and said autonomous exploration could build a more fundamental understanding, with experience eventually feeding shared models. That is a research hope, not a shipping claim. Continuous learning is more likely to complement teleoperation, imitation learning, and simulation than to replace them. Kris said as much. FIRGELLI agrees. Factories that already run teleop cells will not throw those cells away because a kindergarten opened in Shijingshan. They will ask whether the new loop reduces reset rate, improves recovery after a slip, or just adds compute cost next to the same human supervisor.
Shijingshan versus Yizhuang. Method, not just real estate.
Global Times contrasts the kindergarten with existing training bases in Yizhuang and Shijingshan that recreate real-world tasks but lean on human demonstrations, teleoperation, and motion capture. The kindergarten asks a different question: can robots keep adapting from their own experience after deployment. That is a fair research question. It is not proof that the older bases are obsolete, and it is not proof that experience-driven learning already works on a production humanoid. Method claims need hardware that can take the method. If the joints cannot survive wall hits and repeated recovery attempts, the learning loop starves or the robot breaks. Hardware robustness for mistakes is the real bottleneck Kris pointed at. FIRGELLI will keep that sentence in bold when vendors sell continuous learning without a thermal and impact sheet.
The thirty-seven-day build is speed of construction and coordination with city and Shougang Group support. It is not proof that the learning method works. Opening fast is a logistics story. Continuous adaptation after deployment is a controls and durability story. Do not collapse those into one headline. The same caution applies to any soft “first” language around robot kindergartens. Without a clear baseline of prior facilities, prior algorithms, and prior platforms, first claims need caveats. What opened on 1 September is a tactile-plus-experience facility co-branded with Sutton’s team and Tashan. That is enough for the news desk. It is not enough to rewrite how humanoid buyers should buy joints.
What a buyer should ask next.
If you are evaluating platforms that might train here or similar sites, ask the boring questions. Which bodies are on the floor. DoF table and joint map. Rated payload at full arm extension. Continuous-duty torque at a stated temperature for hip, knee, and wrist. How the machine behaves after a hard stop into a wall. Whether tactile skins and joint temperature sensors close a real feedback loop into the learning stack, or sit as brochure sensors. How many minutes of failed attempts a pack and gearbox can absorb before derate. Whether continuous learning is allowed on a production shift or only inside the kindergarten fence. Sutton’s point that environments must be designed for learning is the right engineering instinct. It also means the facility robots are research assets until someone publishes duty numbers.
FIRGELLI’s read after 1–2 September is narrow. A Shijingshan robot kindergarten opened under Tashan Technology and Sutton’s Openmind-linked team. Focus is tactile sensing and experience-driven continuous learning. Public coverage documents the charter, quotes from Sutton, Ma Yang, Kris De Asis, and Wang Peng, a forty-minute spider locomotion demo, and a thirty-seven-day stand-up with city and Shougang support. Missing are humanoid DoF, payload, thermal, and continuous-duty figures for the machines inside. The spider result does not stand in for humanoid continuous learning. Continuous learning complements teleop and imitation; it does not erase them. Hardware that can take mistakes is the gate. Thirty-seven days proves build speed, not learning method. Until Tashan or Openmind publishes a joint map and a measured duty, treat this as a research opening with an honest research gap on the body numbers.
If you are stacking this facility against other Beijing training bases or against compact bipeds on this desk, hold it to the same bar. A named platform. A published force or payload. A measured thermal or runtime limit under failed attempts. Tuesday opened a kindergarten thesis. It did not publish the continuous joints.