Mô hình nền tảng AI và Robotics của Skild

Skild AI is a robotics software company that sells one thing: a robot brain. Instead of writing a separate controller for each machine and each task, it trains a single large neural network, the Skild Brain, intended to run quadrupeds, humanoids, tabletop arms and mobile manipulators. In AI terms this is a robot foundation model: a model pretrained on broad data that can then be adapted, prompted or fine-tuned for many downstream jobs, the way large language models are for text. The company was started in Pittsburgh in 2023 by Deepak Pathak and Abhinav Gupta, both former Carnegie Mellon University professors, as Maginative reported when its first major funding round was announced.

The core problem Skild AI targets is data. Language models learn from trillions of words of internet text, but there is no comparable archive of robot actions. Collecting real robot data means people teleoperating machines one demonstration at a time, which is slow and expensive. Skild AI's July 2025 post argues that many existing robotics foundation models start from a vision-language model, a model trained on images and text, and add a small share of robot data, which it puts at under 1%. Its view is that such models understand objects and words but lack action information and physical common sense, so they mainly show semantic generalization on pick-and-place tasks.

The company's answer is a three-part data recipe. Pretraining uses large-scale simulation and internet video. Post-training adds targeted real-world robot data. Human video is treated as data from one more kind of body, since a person picking up a cup is solving the same physical problem a robot arm faces. In January 2026 Skild AI said it could fine-tune its model to perform new skills by watching videos alone, using less than one hour of robot data. The design is hierarchical: a high-level policy runs at low frequency and decides manipulation and navigation actions, while a low-level policy runs at high frequency and turns those commands into joint angles and motor torques. The deployed model is fine-tuned and distilled, meaning compressed into a smaller network, for each setting.

The defining claim is that the brain is omni-bodied: it does not need to be told what body it is in. In September 2025 Skild AI described training one model across a simulated universe of 100,000 different robots for the equivalent of roughly a thousand years of simulated time. Test robots were held out of training, so every result was zero-shot. When engineers removed a quadruped's lower legs, it adapted in about seven to eight seconds by swinging its thigh joints, while a specialist controller trained for that one robot flipped over. With knees locked in software it walked on three legs within two to three seconds, and with jammed wheels it switched from rolling to walking. The company compares this to in-context learning in language models: the model infers its situation from recent history rather than from new training.

The milestones came quickly. Skild AI announced a $300 million Series A at a $1.5 billion valuation on 9 July 2024, led by Lightspeed Venture Partners, Coatue, SoftBank Group and Bezos Expeditions. It publicly introduced the Skild Brain on 29 July 2025. On 14 January 2026 it announced a $1.4 billion Series C at a valuation above $14 billion, led by SoftBank with NVentures, the venture arm of NVIDIA, among the participants, and said revenue had grown from zero to about $30 million within a few months of 2025. In August 2026 it released S1, a manipulation model that learns tasks from a single video demonstration, and in September 2026 it reported crossing $100 million in annual recurring revenue about ten months after its first commercial deployment, with more than 60 paying customers.

S1 shows how the approach compares with the main alternative, a vision-language-action (VLA) model prompted with text instructions. In Skild AI's own tests after 100,000 hours of pretraining, S1 succeeded on about 66% of unseen tasks when shown one demonstration, against 9% for a language-prompted VLA policy. On seen tasks it reached about 96%. The company estimates that one in-context demonstration is worth about 380 conventional post-training demonstrations, which would take 50 to 100 hours of teleoperation to collect. The comparison against classic robotics is starker still: a hand-engineered controller is tuned for one machine and usually fails when the machine changes, which is exactly the case the omni-bodied tests probe.

The limitations are significant. Almost all published evidence comes from Skild AI's own blog posts and showcase videos, not from peer-reviewed papers or shared public benchmarks, and the September 2025 post discloses no architecture, model size or compute budget. Results such as the 66% unseen-task rate depend on the company's chosen tasks and baselines. Real deployments remain hard: the founders wrote that teaching the model to cook an egg took a week for the first egg and two months before it worked reliably across different eggs and setups. Robot soccer trained by self-play, shown in September 2026, is still far below human level by the company's own account.

In practice, using the Skild Brain is a commercial relationship rather than a download. There are no open weights or public model cards; customers deploy it on robots the company supports and integrates. About 90% of its reported revenue comes from manipulation work and about 10% from mobility, a share that includes autonomous mobile robots from its acquisition of Zebra Technologies' robotics business. Buyers should expect integration work, on-site tuning and the safety engineering any industrial robot requires.