SeeAct AI: embodied AI's endgame is robots that retrain themselves
A Chinese robotics startup founded this year is arguing that the next leap in physical AI will not come from bigger models or more human demonstrations, but from letting robots generate and bank their own experience. Also today: iFlytek's industrial arm ships a small energy model built to run on the factory floor.
Yao Mu, an assistant professor at Shanghai Jiao Tong University, has founded SeeAct AI to bet that embodied intelligence ends in reward-driven self-evolution rather than more training data. In an interview with Leiphone, Mu lays out a three-axis view of scaling: data, model — and experience, the axis he says everyone is under-building. His argument is that human demonstrations only teach a robot what a person would do in a given situation. They do not teach it what happens when its grip lands a centimeter off, or how to recover when a bottle cap will not turn. Those are experiences the machine has to generate itself, be rewarded for, and keep.
The architecture follows from that thesis. Mu's lab has pushed discrete-diffusion vision-language-action models precisely because continuous-diffusion policies — the Pi-family approach — generate smooth trajectories but are awkward to optimize with reinforcement learning, while a token-prediction policy slots into PPO and GRPO more directly. On top of that sits work he describes as converging rather than competing: MM-ACT folding language, video and action generation into one model, and dVLA-RL training across many tasks and scenes instead of one policy per scenario. The leverage he is counting on is cheap experience at scale — a little real hardware plus a lot of interaction inside simulation and world-model environments. He is explicitly relaxed about the sim-to-real gap: if friction is 0.2 in the real world and 0.3 in a simulated one, what the robot learns there is still knowledge, and he cites simulation-trained grasping that beats data-collection farms. The same thesis is visible elsewhere in the field, from Astribot's SmoothRL work on online robot learning to NVIDIA's open-source Cosmos 3 world model.
The commercial claim is the interesting part: SeeAct says it will sell a policy that keeps improving after it ships. The company plans to deliver an edge compute box plus a stack — world model, policy model, reward model and agent system — rather than a frozen controller, and targets roughly 48 hours of autonomous self-improvement for a customer's unfamiliar scenario. If that holds, embodied delivery stops being a project business of twenty engineers for three months per site and becomes something closer to a product. Mu's own account is careful about the state of the art: he puts Generalist's GEN-1.5 at about 58% success on a single in-context demonstration in a scene it has never seen, which he calls academically remarkable and still a long way from 99.99%. He also acknowledges the pushback — critics on Chinese social media have dismissed embodied recursive self-improvement as a repackaged buzzword. The honest read is that the direction is well-argued and the timeline is not yet evidenced: no customer deployment, no success-rate curve, and verification of whether a robot actually got better remains the hardest unsolved piece of the loop.
Antelope shipped industrial model 3.5, and its numbers are about cost, not capability. The iFlytek-affiliated industrial internet company released the model at the 2026 World Manufacturing Convention in Hefei, alongside three agents for compute-power coordination, power trading and zero-carbon industrial parks. Antelope says its time-series foundation model generalizes zero-shot and few-shot across the more than 70% of industrial data that is time-series, forecasts electricity prices at above 85% average accuracy, and tops out at just 100 million parameters — small enough to deploy on edge devices rather than in a data center. The claimed field results are specific: a 5% to 8% cut in blended electricity cost and a 20% gain in storage revenue in pilots for the compute-power agent, and 5% to 10% higher revenue per kilowatt-hour for AI-assisted power trading, which Antelope says has run six consecutive months below the provincial wholesale average price this year. Treat the percentages as the vendor's own. The notable part is the shape of the bet — a domestic industrial stack chasing a 100M-parameter model on the factory floor while the frontier labs chase scale, from the same company that shipped Spark X2.5 on Chinese silicon this month.
What to watch: whether SeeAct publishes a success-rate curve from a real customer site, and whether Antelope's 100M-parameter pitch survives contact with a second year of deployments.
Would you trust a robot whose policy keeps rewriting itself after you bought it — and how would you audit it? Tell us in the comments.
Sources: Leiphone · 36Kr — Intelligent Emergence · Shanghai Jiao Tong University faculty page · RoboTwin (GitHub) · CNR · China Youth Daily · China News Anhui
