Robot Learns New Tasks From Single Video Demo

Industrial robots have traditionally required significant reprogramming whenever a factory layout changes or a new product is introduced. Skild AI has launched a new foundation model called S1 that aims to solve this rigidity by allowing robots to learn previously unseen tasks from a single video demonstration. This approach removes the need for task-specific post-training or weight updates, a technique known as in-context learning that enables immediate adaptation.

The system interprets the intent, objects, and sequence shown in the video and maps them into physical actions without retraining. According to reporting by GN technics/ai (en-US), this capability allows the robot to handle complex, multi-step processes that were not included in its original pretraining data. The goal is to move adaptable robot intelligence from laboratory settings into dynamic manufacturing environments where workflows shift frequently.

Reducing Training Time Drastically

Most industrial robots are built for fixed jobs, meaning each new process requires collecting new data, retraining the model, and running validation tests. S1 changes this workflow by treating a video as a prompt. An operator records the desired task, and the model analyzes the demonstration to understand the required sequence of movements. This allows the robot to execute unfamiliar tasks lasting up to 10 minutes, such as assembling kits or brewing coffee, by composing skills in sequences it has not previously performed.

In testing, Skild AI reported that the time from recording a demonstration to autonomous execution was just 11 minutes in a plant-potting scenario. The model can also adjust if objects move or if the robot encounters errors, recovering without explicit programming. These improvements suggest a significant reduction in the manual labor required to keep robots aligned with changing production needs.

Performance Gains Over Legacy Systems

The efficiency gains are measurable in success rates and training volume. In tests involving new, multi-step tasks, the S1 model succeeded about 66% of the time at each step. By comparison, a similar AI system without this specific in-context learning capability succeeded only 9% of the time, representing a more than sevenfold improvement. This reliability is critical for industrial applications where consistency is required.

Furthermore, showing the robot one short video example was found to be as useful as providing roughly 380 hands-on training examples. For a human operator, collecting those 380 examples manually could take between 50 and 100 hours. By compressing this training data requirement into a single video, the system significantly lowers the barrier to deploying robots for new, specialized tasks.

Deployment in High Precision Assembly

This technology is already being applied in real-world factory settings. Skild AI, NVIDIA, and Foxconn are deploying the Skild Brain on dual-arm manipulators for the high-precision assembly of NVIDIA Blackwell systems. The demonstrated workflow involves installing components like busbars and limit blocks, fastening 16 screws, and adapting to disturbances during the process.

This application requires precise motion, contact-aware control, and the ability to recover when the physical scene differs from the planned sequence. The collaboration leverages NVIDIA’s infrastructure for simulation and training, allowing Skild to generate diverse experiences for its robots. This integration aims to make robot adaptation a standard part of the manufacturing cycle rather than a disruptive event that halts production.