Sanctuary AI Achieves "Zero-Shot" Sim-to-Real Milestone for Hydraulic Dexterous Hands

Vancouver-based Sanctuary AI has announced a significant technical achievement in robotic manipulation, demonstrating "zero-shot" transfer from simulation to reality for its five-fingered hydraulic hands. In a demonstration released today, the company showed its robotic hand repeatedly reorienting a lettered cube—a task notoriously difficult in robotics due to complex contact dynamics—using a control policy that had never been tested on physical hardware prior to the demo.

The Sim-to-Real Hurdle

The achievement focuses on bridging the "reality gap," the physical discrepancy between a digital simulation and the messy, unpredictable real world. While Sanctuary AI previously touted success in using reinforcement learning (RL) to handle unexpected 500g loads, this new milestone indicates that their simulation environments have reached a level of high fidelity where "zero-shot" transfer is possible.

This means the AI policy was trained entirely in a virtual environment and performed seamlessly when deployed on the physical Phoenix robot platform. This stands in contrast to many contemporary approaches that require extensive real-world "fine-tuning" or human-in-the-loop data collection to succeed in contact-rich tasks.

Hydraulic Advantage and Hardware Design

A central pillar of Sanctuary’s strategy remains its commitment to hydraulic actuation, a path less traveled compared to the electric motors favored by competitors like Tesla and Figure. Sanctuary claims that hydraulics provide superior strength, speed, and precision control, which are essential for achieving human-level dexterity. While Sanctuary AI’s demonstration is a significant step for hydraulic hardware, the underlying software approach mirrors a historical breakthrough from OpenAI nearly eight years ago

The proprietary hand used in the demo features a high number of active degrees of freedom (DOF), including finger abduction—the ability to spread fingers apart. This mechanical versatility is increasingly seen as the baseline for "human parity." For comparison:

• Tesla is currently aiming for a 25-actuator "V3" hand to achieve superhuman precision.
• Figure recently teased its 7th-generation hand, which also emphasizes finger abduction and adduction for stabilizing irregular objects.
• Xiaomi has focused on shrinking its hardware to a 1:1 human scale while adding "bionic sweat glands" for thermal management.

The Competitive Landscape of Dexterity

Sanctuary AI’s "zero-shot" success arrives at a time when the industry is debating the best software architecture for manipulation. While Sanctuary relies on robust RL and high-fidelity simulation, other players are taking different routes:

• Sharpa Robotics recently achieved an apple-peeling milestone using a "Mixture-of-Dexterous-Experts" (MoDE-VLA) architecture, though it currently maintains a 30% success rate on that specific task.
• Kyber Labs is distancing itself from monolithic models in favor of a deterministic "primitive workflow" approach for lab automation, prioritizing auditability over the emergent behaviors of RL.
• PSYONIC and NVIDIA are collaborating on real-to-real transfer, using a sensorized prosthetic hand to capture high-fidelity human data for training various robotic embodiments.

By demonstrating that its hydraulic hardware can execute complex in-hand reorientation without real-world training, Sanctuary AI is positioning its "Carbon" AI control system as a leader in generalized dexterity. However, the broader challenge for the Vancouver firm—and the industry at large—remains scaling these specialized skills into a reliable, affordable system capable of handling the vast array of objects found in human environments.