In this conversation, we spoke with Rich Walker, Director at Shadow Robot, about the trade-offs shaping robot dexterity, why reliability often matters more than anthropomorphism in manipulation, and what would need to change for dexterous systems to become dependable infrastructure.
For decades Shadow pursued anatomical precision, yet DEX-EE moved toward fewer fingers and greater structural tolerance; was there a moment during RL training where humanoid fidelity clearly failed, and how did that reshape your long-term positioning in manipulation?
DEX-EE was the result of asking the question “what do you care about more, dexterity+reliability or anthropomorphism”. And DeepMind said they cared more about dexterity+reliability. Our perspective is that if you need to do something that requires the full human ergonomics, then you’ll need an anthropomorphic hand, but if you need dexterity then you can use other morphologies.
In recent DeepMind collaborations, have there been extended training cycles where hardware degradation or unexpected breakage forced a design reset, and what did that reveal about the true durability threshold required for scalable embodied AI?
We spent a number of years abusing the DEX-EE hand to make sure that when DeepMind used it for training they didn’t have these sorts of failures and degradation. The test spec for DEX-EE was actually more brutal than DeepMind really asked for, because we wanted to be confident of its performance. It was very clear to us that it was a systems design issue – both the raw mechanical performance and the control stack on top of that – to get the necessary robustness.
Shadow has remained relatively small and grant-backed while others scaled through venture capital; since 2022, has there been a deployment or partnership opportunity you could not pursue at full speed, and how do you evaluate that trade-off today?
We’ve always used grant funding for what it’s good for – speculative collaborations across organisations – and sold product to fund our internal development. We work with customers who have a variety of needs around development, and for some of them they have access to the necessary resources to pay for the development directly; others use us as a springboard to get to their Series A funding.
That said, we’re looking with our ARIA Robot Dexterity programme work at some bigger opportunities that might well need significant capital to scale.
With high-density tactile sensing embedded in DEX-EE, have you seen situations where sensor consistency materially accelerated learning compared to vision-only setups, and do you believe tactile data becomes a strategic control layer by the end of this decade?
I can say that there are plenty of tasks that humans do without vision, exploiting tactile sensing to do them, and there is no reason to assume robots won’t have the same processes.
Reducing degrees of freedom can look like simplification, yet it often requires sacrificing elegance; was there internal resistance to abandoning five-finger symmetry, and does subtractive engineering now define Shadow’s competitive philosophy?
DEX-EE was actually more complex in many ways than the classic Shadow Hand – it uses 5 motors to drive 4 joints. If we had to explain our design philosophy, it would be to observe that over almost 30 years we’ve built a vast range of hand-like things, so we’ve explored most of the trade-offs “in practise” rather than purely on paper, and that helps us generate the right hardware for customers more easily.
If OGRES allows engineers to generate optimized end-effectors through simulation rather than physical iteration, where does Shadow’s core advantage sit in that future, and does that shift your role from hardware maker to protocol setter?
That “if” is doing a lot of work. The ambition for OGRES is certainly that it will significantly simplify the design of usable robot hands, so maybe we’re looking more at the way that good PCB layout software can verify that the design meets manufacturing and functional requirements – it’s still up to the circuit designer to architect the design and select the key components. And that expertise in architecting designs and working with both low and high TRL components is something we have in depth.
In testing high-torque compact joints, have you encountered physical ceilings in motor density or heat dissipation that fundamentally limit current architectures, and how far are you willing to depart from conventional electric drives to break that ceiling before 2030?
We started out building robots using braided pneumatic actuators (we called them Air Muscles) and we transitioned across to electric motors in the 2006-2008 period, because we were able to get technical advantages from electric rather than pneumatic systems. Transitioning to alternate actuation technologies when they offer superior performance is part of that architecting the design process.
Lower-cost dexterous hands are entering the market rapidly; have you lost real contracts where cost outweighed durability or data quality, and what layer of value do you believe remains structurally non-commoditizable?
We’re actually very pleased that there are a number of low-cost robot hands out there. It has opened up a supply chain in ways that weren’t available 10 years ago. It also helps us answer the question from potential customers about “can you make a hand for my budget of £x”.
We work with customers in the long term, and they come back to us for continued engagement and support 5 or 10 years after first buying hardware.
The 2019 cube-solving milestone proved feasibility but did not immediately translate into mass deployment; from your vantage point in 2026, is the core friction now hardware survivability, algorithmic abstraction, or integration complexity?
That was a fantastic demonstration of how to use a new approach to Reinforcement Learning to solve a problem. Like any “state of the art” game-changing approach, it took a while for people to work out what you could do with the technology. You’ll notice that it went from “you need a 6144 core cluster” to “you can run it as a ‘hello robot’ example in ISAAC SIM” – showing the way the cutting edge becomes quietly normalised.
I would say that the challenge now is around controlled application of force to objects under manipulation. To date, no-one seems to have cracked that.
In elderly assistance scenarios such as bed transfer or object retrieval, have there been moments where anthropomorphic expectations conflicted with your engineering philosophy, and how do you balance mechanical clarity with emotional comfort in domestic environments?
When we’ve worked with researchers looking at assistive scenarios, we’ve always been part of a wider discussion around ethical adoption of robot hardware. Not only does the hardware have to be safe, it has to be acceptable to the people who will co-exist with it. If you don’t get both of those parts right, you’re not doing the basic engineering!
Working within European funding and regulatory frameworks, have policy structures ever slowed iteration cycles, and do you see Europe’s strategic edge emerging more from standards influence than manufacturing scale?
A lot of people miss the fundamental goal of the European political project – which is (in an over-simplified way) to ensure that no-one goes across the Rhine with “weapons hot” again. Part of achieving that is of course to create a stronger economic cohesion with excellent organisations that are long-term stable employers and contribute to both individual and societal needs.
European regulatory frameworks should be seen with that view, as should European funding models – which typically are long-term substantial contracts won by multi-national teams representing research and industry, and are aligned with long term societal challenges rather than short term market needs.
I would argue that the strategic edge in Europe is this societal view, rather than a purely market or political view as in other regions.
And manufacturing scale is something Europe does do when it wants to. There’s a reason so many of the classic “hidden champions” are European companies.
If manipulation becomes dependable infrastructure by 2030, what concrete milestone would signal that shift to you, and where does Shadow intend to sit in that ecosystem when dexterity is no longer experimental but assumed?
For a milestone that manipulation is dependable, I’d expect to see automotive companies deploying it at scale across manufacturing facilities in the same way as today they deploy welding and riveting robots.
Our role at Shadow Robot is to bring deep expertise in the design and development of robot dexterity and applications of robot dexterity. In the same way as there isn’t one car manufacturer, there won’t be one robot hand.
