Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Physical Self-Play is a research result from Skild AI showing that its S1 robotics foundation model can be post-trained with self-play reinforcement learning. The policy was given a single objective, to score, and trained against recent versions of itself in NVIDIA's Isaac Sim. S1 is described as a model that learns tasks from in-context demonstrations and is pre-trained on human data.
Skild reports over 140 years of simulated play. Behaviors such as dribbling past a defender, shielding the ball and tackling emerged without hand-crafted rewards for them. Early in training the policy could barely walk, and it later learned to get back up after falling. The trained policy was then transferred to a humanoid robot and put through a match.
This is a blog announcement of a research result, not a released product. No public model, code, weights, paper or API is listed. Skild says it plans to extend the approach to other everyday robotics tasks and to larger teams in a future release. The sim-to-real claims come from the vendor's own post.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Humanoid robot model S1 trained via self-play reinforcement learning in Isaac Sim over 140+ years of soccer with an automatic training ladder, then deployed to real robots.
Closest neighbours by meaning, not just by category.
Model · Skild AI
Skild AI's robotic foundation model that performs new tasks from a single video demonstration, with no fine-tuning.
Research · Tsinghua University
Tsinghua research system in which a humanoid robot learns to play badminton rallies from limited human motion data.
Model · Generalist AI
Robot foundation model from Generalist AI that learns new tasks in-context from a short demonstration.
Model
27B vision-language robot policy trained only in simulation, run zero-shot on a real bimanual robot.
Hardware · Roboparty
Open-source full-stack humanoid robot platform from Roboparty: hardware designs, firmware, control and learning code.
Research
Research system that transfers real tennis players' motions onto a Unitree G1 humanoid for rallies and serves.
