Sanctuary AI shows how reinforcement learning can control hydraulic robotic hands

Sanctuary AI this week released a video demonstrating advanced manipulation skills with its hydraulic robotic hands. The company said it used dexterous policies to train the hand, which features strength, speed, and a high degree of freedom.

Many companies have successfully used reinforcement learning in simulation to learn complex policies for quadruped robots. In contrast to locomotion, training and transferring dexterous manipulation policies with five-fingered hands have found limited success.

The intricacies of robotic hands make them challenging to code, Simulation and reinforcement learning (RL) techniques offer a clearer path to achieving industrially relevant behaviors for manipulation, according to Sanctuary.

In the video below, the company demonstrates an in-hand reorientation policy trained in simulation being executed in the real world and against gravity.

The training session objective was to create a policy that enables the system to autonomously achieve a goal such as grasping and turning a cylinder without dropping it. Sanctuary said its proprietary RL approach has enabled in-hand reorientation under an extreme disturbance — a 500 g (17.6 oz.) load that was not encountered during training.

What makes this case different is the strong and dexterous hand hardware serves as the vehicle for effectively conducting dexterous, physical work, asserted the Vancouver, Canada-based company.

Sanctuary collaborates with NVIDIA

Sanctuary AI noted that its proprietary robotic hands have 21 active degrees of freedom (DoF), which enables finger abduction and advanced in-hand manipulation. It also said hydraulic actuation offers strength, speed, and control, while the system’s compact hydraulic valves offer a promising path to achieving human-level dexterity.

By contrast, Boston Dynamics used hydraulic actuation in its previous Atlas models but switched to electric actuation when it decided to create a commercial model.

The company uses NVIDIA Isaac Lab to simulate dexterity-focused training environments. Isaac Lab is an open-source, unified framework that enables the training of robot policies with high-fidelity simulation.

Built on NVIDIA Isaac Sim, Isaac Lab uses PhysX for physics simulation and RTX rendering to bridge the gap between simulation and perception-based robot training. This can help researchers and developers build autonomous robots more efficiently, according to Sanctuary AI.

Founded in 2018, Sanctuary Cognitive Systems Corp. has been recognized as a leader in intellectual property around general-purpose robots and embodied artificial intelligence. Morgan Stanley recently ranked it third globally for published U.S. patents.

Sanctuary unveiled the latest iteration of its Phoenix robots last year and raised funding, bringing its total to $140 million.

---

---