Latest23 Sep 2026OtherResearch
Physical Self-Play
A post-training method, not a new model: S1 plays against recent copies of itself in NVIDIA Isaac Sim with one objective, scoring at robot soccer. Over about 140 years of simulated play, dribbling, shielding and tackling emerged without hand-written rewards, and the policy transferred to a real humanoid. Skild pitches it as a way past the human-level cap of training on human data.
- Method applied to the S1 model
- Trained in NVIDIA Isaac Sim
- About 140 years of simulated play
- Transferred from simulation to a humanoid robot (model not named)
18 Aug 2026Robot model (AI)Released
Skild S1
Skild's flagship robot foundation model, built as an in-context learner: shown a video demonstration, even one recorded on a different body, it carries out the task on its own robot without fine-tuning or post-training. Skild says it handles seen and unseen, short and long tasks of up to ten minutes and is already at work with commercial partners; early access is by sign-up.
- Learns tasks from a video demonstration in context, no fine-tuning
- 96% accuracy on seen tasks at scale (company figure)
- One in-context demonstration worth roughly 380 post-training examples on unseen tasks
- Long-horizon tasks up to ten minutes
- Specific robot bodies, parameter count and open weights not stated
12 Jan 2026OtherResearch
Omni-bodied learning from human video
Research write-up on what Skild calls omni-bodied learning: its model is fine-tuned to perform new skills by watching human videos, plus less than an hour of robot data. The goal is to tap the millions of instructional videos online instead of costly teleoperation. The post names no model version and does not say which robot hardware was used.
- New skills from human video plus under 1 hour of robot data
- No model version named
29 Jul 2025Robot model (AI)Released
Skild Brain
Skild's original robot foundation model, introduced as a single omni-bodied brain for quadrupeds, humanoids, table-top arms and mobile manipulators. A low-frequency planning policy feeds a high-frequency controller that sets joint angles and motor torques. In March 2026 ABB Robotics and Universal Robots agreed to integrate it into their robots, and it is being deployed on dual-arm manipulators with NVIDIA and Foxconn.
- Embodiments named: quadrupeds, humanoids, table-top arms, mobile manipulators
- Hierarchical: low-frequency manipulation/navigation policy plus high-frequency low-level controller
- Being integrated into ABB Robotics and Universal Robots portfolios (Mar 2026)