Sanctuary AI Touts Reinforcement Learning Success for Dexterous Robot Hand Manipulation

Image credit: Sanctuary AI

Sanctuary AI Claims Dexterous Manipulation Breakthrough with Reinforcement Learning

Sanctuary AI, the Vancouver-based company developing the humanoid robot Phoenix, has announced a notable advancement in robotic dexterity. The company demonstrated its five-fingered robotic hand performing in-hand object reorientation using a policy trained via reinforcement learning (RL) in simulation.

Watch the demo above

Sim-to-Real Dexterity Under Pressure

While RL has seen considerable success in teaching legged robots to walk and run in simulation before transferring those skills to physical hardware, achieving similar results for complex, multi-fingered manipulation has proven more challenging. Tasks involving the fine motor skills of a hand are sensitive to subtle physical differences between simulation and reality, often hindering successful transfer.

Sanctuary AI states its proprietary RL approach enabled a physical robot hand to reorient an object even when subjected to an unexpected 500g load – a disturbance significantly different from its training conditions. The company describes this demonstration, involving transferring the learned policy from simulation to their physical hardware under duress, as a significant step forward for the field.

Hardware Matters: Hydraulic Hands

A key component of this achievement, according to Sanctuary AI, is the design of their robotic hand. Unlike many robotic grippers focused primarily on grasping, Sanctuary's hand emphasizes dexterity. It features a high number of active degrees of freedom, including finger abduction (the ability to spread fingers apart), which allows for more complex in-hand manipulation.

The hand utilizes hydraulic actuation, a less common approach in dexterous robotic hands compared to electric motors. Sanctuary AI claims this choice provides superior strength, speed, and control precision. They also highlight their compact hydraulic valve technology as a pathway towards achieving human-like dexterity in robots.

Image credit: Sanctuary AI

Context and Ambition

Founded in 2018, Sanctuary AI aims to create general-purpose robots capable of performing a wide range of tasks, positioning its technology as a solution to labor shortages. The Phoenix robot, unveiled in 2022, serves as the platform for integrating the company's AI control system, known as Carbon, and its dexterous hands.

This demonstration represents progress in developing capable manipulation systems, a critical component for general-purpose robots intended to work in human environments. While the company highlights the unique capabilities of its hardware and RL techniques, the broader challenge remains in scaling these abilities to handle a vast array of objects and tasks reliably and affordably. The company positions this demonstration as evidence of its leadership in the field, citing its intellectual property portfolio and focus on dexterous intelligence.

Read the original Sanctuary AI blog post here: https://www.sanctuary.ai/blog/sanctuary-ai-controlling-advanced-hydraulic-hands