The Final Frontier of Physical AI: How Sanctuary AI is Disrupting Dexterity with Hydraulics and Deepening Japanese Alliances
- Miki Sadinov
- Jul 7
- 5 min read
Humanoid Summit2026
「Dexterity: The Final Frontier for Physical AI」
James Wells, CEO at Sanctuary AI
While the global artificial intelligence race has seen billions of dollars poured into compute, foundational models, and data pipelines, the physical application of AI has remained stalled at a single, frustrating bottleneck: dexterity. Robotic arms can move, and AI brains can think, but the ability of a mechanical hand to manipulate the world with human-like precision has remained the "number one bottleneck" preventing widespread industrial automation.
At the first annual Asia Summit in Tokyo, James Wells, the co-founder and CEO of Canadian robotics pioneer Sanctuary AI, took the stage to present a radical path forward. Entitled “Dexterity, the final frontier for physical AI,” Wells’ presentation laid out how Sanctuary AI is combining state-of-the-art software with a revolutionary miniature hydraulic-actuated multi-finger hand to unlock the future of physical labor.
A Three-Generation Legacy: Announcing the ZEON Corporation Alliance
For Wells, returning to Tokyo to announce Sanctuary AI’s latest breakthrough was not simply a business milestone, but a deeply personal homecoming. Wells revealed that his family’s ties to Japanese industry span three generations:
A Deep Historical Connection: During the 1960s, 70s, and 80s, Wells’ grandfather and father worked closely with Japan's premier trading firms—including Nissho Iwai, Mitsui, and Mitsubishi—selling Canadian lumber. Wells noted the profound honor he felt returning to Tokyo, representing the third generation of his family's business relationship with Japan.
The ZEON Corporation Partnership: Sanctuary AI officially announced a landmark joint development partnership with Japan’s ZEON Corporation. The alliance is specifically tasked with co-defining and co-developing "new materials" optimized for dexterous robotic manipulation. Wells emphasized that this advanced material-science collaboration represents a major competitive differentiator, creating a technological moat that global competitors cannot easily replicate.
Solving the "Both" Dilemma: A Task-First AI Platform
A long-standing debate within the robotics community has pitted simple, low-cost off-the-shelf grippers against complex, multi-finger anthropomorphic hands. Sanctuary AI’s research, spanning discussions with hundreds of customer companies, has yielded a definitive conclusion: the industry needs both.
Off-the-Shelf Grippers: By injecting advanced AI policies into simple parallel or suction grippers, Sanctuary AI can deploy robots immediately to deliver rapid, direct return on investment (ROI) for basic industrial tasks.
Multi-Finger Anthropomorphic Hands: To unlock the long-term goal of human-like versatility and automate highly complex, non-repetitive tasks of the future, the company is parallelly developing five-finger robotic hands.
A Unified AI Platform: Rather than maintaining fragmented systems, Sanctuary AI operates a single AI platform capable of seamlessly controlling suction, parallel grippers, and fully anthropomorphic hands alike. This "task-first" philosophy ensures that the software adapts dynamically to the physical tools required for any given job.
Real-World Proof: Sub-Millimeter Insertion and 0.1mm Tolerances
Moving past sterile lab demonstrations, Wells showcased video evidence of Sanctuary AI robots operating directly on active automotive manufacturing lines under strict, unforgiving industrial cycle times:
1. High-Precision Sheet Metal Positioning
In an automotive plant, Sanctuary's AI system was tasked with picking up randomized, irregularly shaped sheet metal plates and placing them onto a welding jig where nuts and bolts are subsequently welded.
The Engineering Challenge: The system must seamlessly blend the pick-up and placement phases into a single, continuous, fluid motion.
0.1mm Insertion Tolerance: To secure the sheet metal, the hand must thread the plate onto tiny guide pins with an incredibly tight 0.1 mm allowable tolerance. Traditional physical programming and fixed automation struggle to exceed a 1-to-2 mm margin of error.
Adaptive Motion Planning: Sanctuary's AI policy does not rely on rigid, repetitive trajectories. Instead, it dynamically senses how the metal was initially grabbed (compensating for micro-slippages) and adapts its path in real-time to slide the metal onto the pins perfectly.
2. Wire Harness Plugging on Moving Conveyors
Deemed one of the "worst tasks" in electronics and automotive assembly due to its flexible, non-rigid nature, wire harness plugging has historically baffled roboticists. Wires bend, deform, and behave unpredictably.
Constantly Moving Target: Sanctuary's robot had to grasp a flexible wire plug and insert it into a sub-millimeter port sitting on a constantly moving conveyor belt.
99.5% Success Rate: Utilizing custom end-effectors guided by dynamic vision-based AI policies, the robot achieved an extraordinary 99.5% success rate while fully maintaining the plant's high-speed industrial cycle times.
Extreme Sample Efficiency: 5 Hours of Data, 12 GPU Hours
The true commercial viability of physical AI hinges on its ability to handle unexpected real-world errors and adapt on-site without millions of dollars in training costs.
Autonomous Error Recovery: In a "blooper reel" shared by Wells, when a wire harness plug deflected or failed to insert on the first try, the AI policy immediately recognized the failure. Without halting the production line, the robot adjusted its grip and successfully re-inserted the plug. Furthermore, it accelerated its movements to process the temporary backlog of oncoming plugs.
Next-Generation Sample Efficiency: Traditional imitation or reinforcement learning models require months of data collection and massive compute budgets. Sanctuary AI’s novel architecture achieved this level of precision with less than five hours of on-site teleoperation data.
Ultra-Low Compute Training: The resulting AI policy was fully trained using less than 12 GPU hours. This unprecedented efficiency allows Sanctuary AI to deploy hardware to a customer's facility and fine-tune highly specialized industrial AI policies directly on-site in a matter of hours.
The Miniature Hydraulic Hand: Replicating Human Power Density
Sanctuary AI began its journey into five-finger hand development about four and a half years ago. After experimenting with electromechanical (motor-driven) and cable-driven designs—which repeatedly failed due to joint wear, lack of durability, and inadequate precision—the company turned to hydraulics.
By developing custom, highly miniaturized pistons and actuators, Sanctuary AI has engineered a five-finger hydraulic hand with profound physical advantages:
20x Power Density: Reaches a force output 20 times higher than electromechanical hands of the same size, packaging immense strength into the confined space of a human-sized palm and fingers.
Instantaneous Response Times.
Submillimeter Fingertip Precision.
Dramatically Reduced Thermal Limitations: Eliminates the severe overheating issues common to electric motors running continuously under heavy loads.
The Unedited, Single-Cut Demonstration: Wells presented an unedited, single-cut video showing a teleoperator executing a complex assembly. The robot picked up a small metal bracket, aligned it, and threaded a nut onto a bolt using its fingertips. This was made possible by high-fidelity tactile sensors lining the fingers, which transmit real-time physical feedback (pressure, friction, threading resistance) back to the operator. This seamless human-in-the-loop operation also serves as a pipeline for gathering pristine, high-quality training data for future autonomous policies.
Heavy Tool Manipulation: Because of its hydraulic strength, the hand can easily support heavy, commercial hand tools (such as an electric drill driver) while maintaining delicate finger dexterity. The robot can securely grip a drill, modulate its trigger pull, and align the rotating drill bit precisely with a tiny screw head.
The Shovel Analogy: A Disruptive Shift in Actuation
To frame the significance of this hydraulic breakthrough, Wells drew a historical parallel from Clayton Christensen’s seminal book, The Innovator's Dilemma.
In the mid-20th century, the global excavation and construction market was entirely dominated by massive cable-and-pulley shovel manufacturers. When fledgling hydraulic technology first emerged, it was dismissed as a niche, unproven novelty. Yet, within just 10 years, the superior power density, responsiveness, and reliability of hydraulics captured 100% of the market share, completely wiping out the legacy mechanical shovel manufacturers.
Wells concluded with a bold prediction: "We believe we are currently at the exact same disruptive inflection point in the history of robotic hand actuation, and that hydraulics will trigger a paradigm shift."
