Skild AI launches robot model trained from one video

Skild AI has launched S1, a robot foundation model that learns new tasks from a single video demonstration. The system was developed using NVIDIA AI infrastructure.

S1 is designed for industrial settings where workflows, layouts and product lines change regularly, and robots often need new programming for each variation. According to Skild, the model uses a video prompt to interpret a task and carry it out without changing its weights or undergoing task-specific post-training.

An operator can record a video of a job and provide it to the model as an example. The system then identifies the intended sequence, the objects involved and the actions required for the robot on site.

This differs from many industrial robots trained for fixed routines, which usually require more data collection, retraining and testing when a new process is introduced. Skild said S1 can handle previously unseen tasks lasting up to 10 minutes.

Examples include plant potting, pancake making, pour-over coffee brewing and kit assembly. These jobs can involve dozens of manipulation steps and require the robot to combine skills in a sequence it has not performed before.

In one plant-potting test, Skild said it took 11 minutes to go from recording a demonstration to autonomous execution on hardware. The company added that the model can adjust when objects are moved, recover from errors and continue through multistep tasks when conditions differ from the original setup.

Skild published benchmark figures alongside the launch. In tests on new multistep tasks, S1 succeeded about 66% of the time at each step, compared with 9% for a similar AI system.

The company also estimated that one short video example can be as useful as roughly 380 hands-on training examples. Collecting that volume manually could take 50 to 100 hours, it said.

Commercial push

The launch comes as Skild says it has reached a USD $100 million annual revenue run rate within 10 months of its first commercial deployment. It also says it has built more than 60 deployment partnerships across manufacturing, logistics, inspection, security and food preparation.

Deepak Pathak, Co-Founder and Chief Executive Officer of Skild AI, described the company's view of the shift in robotics. "Learning by experience, and not preprogramming, is the step change that has happened in robotics," Pathak said.

He also linked that progress to NVIDIA's software stack. "NVIDIA Isaac Lab and NVIDIA Cosmos technologies help Skild create the scalable, diverse experience its robots need to learn across many scenarios and embodiments," he said.

Skild said data from commercial deployments can contribute to the broader model when customer agreements allow. The company argues that this could reduce the need to build a new dataset and training run each time a customer modifies a task or production environment.

Factory deployment

That work is already being applied in electronics assembly. Skild, NVIDIA and Foxconn are deploying the Skild Brain on dual-arm manipulators for high-precision assembly of NVIDIA Blackwell systems.

In one workflow described by the companies, a robot installs a busbar and limit block, fastens 16 screws and adapts to disturbances during a multistep task. Skild said the work requires precise motion, contact-aware control, sequence tracking and recovery when the scene differs from the original plan.

Shared tools

The collaboration spans synthetic data generation, simulation, model training and deployment. NVIDIA said its Cosmos models are being used to diversify training data and convert video into structured descriptions, while Cosmos Curator helps annotate, filter and organise data.

Skild is also using NVIDIA Omniverse libraries, Isaac Sim and Isaac Lab to train and validate the model before deployment in physical settings. The company said the Newton physics engine helps its engineers model forces, contact, collision and pressure more accurately in simulation.

The two groups are jointly developing GPU-accelerated simulation solvers for modelling how robots touch, grip and manipulate solid objects. NVIDIA said those solvers will be made available to developers as part of Newton.

For later stages of development, Skild is using NVIDIA Nsight tools to identify performance bottlenecks during training and TensorRT to improve inference speed for robot responses in the field. The companies said this links data, simulation, training and deployment in a single development process.