One of the most complex skills for a robot is not picking an object up, but holding and reorienting it as needed. Humans do this without a second thought: we pick up a block, turn it in our fingers, and place it on the desired side. For robotics, this remains a non-trivial task. That's precisely why the new demonstration from Sanctuary AI is noteworthy.
What Happened
The Sanctuary AI team released a video showing their proprietary hydraulic hand autonomously manipulating a letter block – continuously reorienting it while in its grasp. This is known as in-hand manipulation, or manipulation “within the hand,” meaning it is done without setting the object down and re-grasping it.
The key point is that the system operates in a zero-shot mode. Simply put, the robot completes the task without any prior training on this specific object. There are no pre-recorded movements, no special data labeling for this particular block – the hand adapts in real time.
Why It's Difficult
To grasp the scale of this challenge, just imagine what happens when you turn an object in your hand. The fingers must constantly adjust their pressure, track the object's position, and predict how it will shift with the next movement. All this occurs under conditions of uncertainty: the object might be slippery, angular, or have an uneven weight distribution.
Hydraulics add another layer of complexity. Unlike electric actuators, hydraulic systems offer more power and flexibility, but controlling them with precision is harder. Sanctuary AI has long been betting on hydraulics in its designs, considering it a promising technology for tasks that require delicate physical contact.
“Zero-Shot Learning” – Not Just a Buzzword
In the world of machine learning, zero-shot learning refers to a system's ability to handle new situations it has not encountered during its training. This is especially valuable for object manipulation: in a real-world environment, a robot will have to deal with thousands of different objects, and training it on each one individually is unrealistic.
If this approach proves to be robust, it means the hand can generalize its knowledge of the physics of object interaction – and apply it to unfamiliar objects. This is a qualitative leap compared to how manipulation and grasping tasks were previously solved in robotics.
Context: Where the Field is Headed
Object manipulation is one of the key challenges the entire humanoid and service robot industry is trying to solve. While most robots can carry things or press buttons, fine motor skills remain a weak point. A warehouse robot that drops fragile items, or a manipulator that needs an hour of calibration for a new part, is not a viable solution for the real world.
In this context, Sanctuary AI is forging its own path: the company is focused on creating a general-purpose robotic platform capable of performing diverse tasks in industrial and logistics environments. The block demonstration is not an end in itself, but an illustration of a fundamental skill, without which most practical applications would be impossible.
What Remains Unknown
A video demonstration is not the same as a documented, reproducible result. It's still unknown how reliably the system performs on objects of different shapes, weights, and surface textures. There is also no public data on how quickly its performance degrades when handling more complex objects.
This doesn't mean the result is questionable – it's just that any demonstration shows a best-case scenario, not an average one. The team's next steps will show how well the “zero-shot learning” approach scales in practice.
In any case, the direction is clear: robots are learning to hold things in their hands just like humans do. And judging by the pace at which similar demonstrations are being released by various teams, this task is moving from the “nearly impossible” category to the “solvable in the coming years” category.
