Brett Adcock (Figure) on the Firs... | Over The Horizon Summary

1. Key Themes

Fleet Fungibility as a Core Infrastructure Principle

Figure's approach to continuous operation isn't just about one robot running longer — it's about building a self-managing fleet. Robots communicate autonomously, swap positions during low battery, and route themselves to maintenance. Any robot in the fleet can slot into any role.

"We have hundreds of robots in the office. They can just call robots in the fleet and they'll just come in and dock. They can dock anywhere. And so, you know, in some way we could basically just run for months and months and be like subbing out different robots, robots having problems." 00:03:21

Data Is the True Rate Limiter for Humanoid AI — Not Compute, Not Hardware

Brett is explicit and emphatic: if Figure had the right data today, general robotics would be effectively solved. Compute and manufacturing are secondary constraints.

"If we had the right data we could snap our fingers I think general robotics would be solved today. I think we'd be able to do most things a human can in the world... right now our biggest constraint is data." 00:25:56

Transfer Learning Across Diverse Tasks Is a Structural Moat for Humanoids

A single humanoid platform accumulates benefits from every new task it learns. Non-humanoid robots cannot leverage this cross-task transfer. Brett gave a vivid empirical example: adding non-fridge manipulation data dramatically improved fridge task performance.

"We went from like 60 success rate with fridge only data and then we train that same model again with more data... not even increasing the fridge data and the numbers went up. And that's when we were like holy cow like the this there's a real benefit to having like extremely diverse like data." 00:13:58

"The unfair advantage is like a humanoid that gets good at one thing say logistics — every robot in the fleet will get better at logistics the same level performance, which is unlike humans." 00:13:58

Figure Four Is the "iPhone One Moment" for Humanoid Robotics

Brett describes Figure 4 as unrecognizable compared to prior versions, designed from the ground up around Helix 3 and the data problem. He explicitly frames it as the architecture that will scale to millions of robots.

"Figure four will be the largest step up we've ever made between versions in history... it'd be like an iPhone one moment for the space it is so dramatically different on both on every aspect... it's unrecognizable as a robot." 00:18:27

"We've designed figure four for Helix, Helix three was designed for data, and the whole robot was designed around data which is our like the largest limiting factor by far." 00:18:27

The Physical World Is a Human Operating System — Humanoid Form Is Not Optional

Brett makes a philosophical but practically grounded argument: the built world was optimized for an average human body. Any robot that deviates from that form cannot fully interact with it. This is his core reasoning for rejecting "go beyond human form" arguments.

"Eight billion humans got dropped on here today and then we work to go build a world around it... we live in a human operating system in the physical world and we built that human operating system so we can interact with it every single day. There is no greater general purpose machine than a human just because of that very reason." 00:39:02

Hand Parity With Humans Is Prerequisite for AI Learning at Scale

The hand is not merely a mechanical end-effector — it is a data collection surface. A hand that cannot replicate human kinematics cannot learn from human data. Brett argues the new Figure 4 hand has more actuators than the rest of the robot's body combined, and that tendon-based hands are a dead end.

"In order to get to human-like intelligence having a hand at parity — a robot hand at parity with human hands — is fundamental... our current hand is not doesn't have the same kinematic mapping as a human... if we want to solve general purpose this in the robot, a human hand at parity with a human hand is critical." 00:35:20

Generalization Is the Only Goal That Matters in 2026

Brett is unambiguous that manufacturing, supply chain, and unit economics are downstream problems. The entire company is organized around solving generalization first.

"In 2026 figure needs to solve generalization... that's the number one goal by a mile. It's not manufacturing, it's not supply chain, it's not any of these things. It's generalization." 00:55:06

On-Device AI Is Non-Negotiable for Real-World Deployment

All Helix 2 inference runs entirely on-board with no network dependency. Brett frames this as critical for speed, reliability, and the ability to operate in degraded network environments.

"The brains are in the torso chest right of the robot and there's all that locally on board with no — there's like no network needed to do any AI inference." 00:05:43

"We are largely memory constrained on device for inference today. We do have a planned upgrade coming on the robots for future generation to help with this." 00:12:30

China Supply Chain De-Risking Is Essentially Complete

Despite designing nearly everything in-house — motors, gearboxes, PCBs, cameras, sensors — Figure has moved almost all manufacturing outside China, quietly and ahead of most of the industry.

"My forecast is maybe by next quarter I don't think we'll have anything coming out of China on the supply chain side... we've moved most of our supply chain outside of China a lot of last year for tariffs, geopolitical risks, everything else." 00:24:06

Hark — A Separate AI Lab Within Figure — Is Building Next-Generation Models and Voice

Brett quietly revealed a 70-person internal AI lab called Hark, started last summer, working on real-time speech-to-speech and next-generation AI models. Every Figure robot already runs the Hark voice model.

"I started a new lab called Hark last summer. We have a team about 70. We're working on like new kind of AI, next generation AI models and some hardware, and one of the areas we're spending a lot of time is real-time speech to speech. The robots — every robot Figure now has the Hark voice model on it." 00:49:49

2. Contrarian Perspectives

Tendon-Based Hands Are a Local Maximum and the Wrong Path for Humanoids

The industry has pursued tendon-based hands for years. Brett designed and then killed his own tendon-based hand program at Figure after a month of firsthand experience, calling it an engineering mistake.

"Tendons is for sure the wrong way to go with hands. It's a local maximum and it is not the right engineering design long term for humanoids and I learned that firsthand." 00:32:40

A Single Unified Neural Network Will Beat Specialized Task Models — No "App Store" Needed

The dominant industry assumption is that specialized models per task will be required. Brett argues the opposite: one large model trained on all diverse data outperforms any ensemble of specialized models, and the app store model is fundamentally wrong for humanoids.

"I don't think the — one beauty of having a humanoid robot is you can basically benefit from transfer learning. You can have a robot that can have one single ideally one single neural network that is pulling in all this very diverse different types of use case work... this is what you don't get from having multiple different robot form factors." 00:13:39

Going Beyond the Human Form Factor Is a Mistake, Not an Advantage

Many roboticists argue that the humanoid form is a constraint and that future robots should exceed human physical configurations. Brett disagrees entirely, grounding the argument in world infrastructure design.

"There is no getting better like you know what I mean. It's just like it is. We've designed it specifically for a 10-fingered human that's roughly five foot two, five foot five that can do everything in the world... if you're two feet tall you can't reach up on the cabinets, and if you're 20 feet tall I obviously won't be able to work well in the world either." 00:39:02

Most Data Problems Are Misdiagnosed as Compute Problems

The industry tends to frame AI capability limits as compute constraints. Brett says this is wrong in the near term — the right data would solve general robotics today, with the hardware and compute already sufficient.

"If we had the right data we could snap our fingers I think general robotics would be solved today... right now what we need to solve is a general purpose machine that has common sense reasoning like a human... we're not compute constrained — we have a great relationship with NVIDIA and they've supplied us we just launched a new cluster of B200s." 00:25:56

Robots Should Coordinate Entirely Through Vision, Not Explicit Network Messaging

In the bed-making demo, two Figure robots coordinated complex collaborative tasks (like timing sheet tensioning) purely through visual head-nod gestures — no explicit inter-robot messaging. This runs counter to the assumption that robot collaboration requires dedicated communication protocols.

"In this case there's no explicit messaging between the robots. They like — they're coordinating their actions fully visually with the head nods. So they're not — that's how they're communicating." 00:44:00

3. Companies Identified

Figure

AI robotics company building general-purpose humanoid robots. The subject of the entire episode — running the first documented 8-hour fully autonomous commercial shift, deploying Helix 2 end-to-end neural network policies, operating an internal AI lab (Hark), completing critical design review for Figure 4, and training on a new B200 cluster.

"We just finished our critical design review for figure four so it's our last major design gate as we like enter design lock and ship parts out." 00:18:27

NVIDIA

GPU and AI compute partner to Figure. Supplied compute infrastructure including the new Blackwell B200 cluster that went live during the episode's recording month.

"We just launched a new cluster of B200s that we announced a few weeks back that just went live this month. We're training in the pre-training some of the largest models we've ever done." 00:25:56

4. People Identified

Brett Adcock

Founder and CEO of Figure. Former founder of Vettery (acquired by Adecco) and Archer Aviation. Described as working 80-100 hour weeks, present at the office 7 days a week until midnight, and personally present for critical hardware decisions including hand design reviews. Running one of the most ambitious humanoid robotics programs in the world.

"We're working 80-100 hour weeks here. Just like how do we go solve for a general purpose machine — that's what we're pushing now." 00:30:06

"I personally was there when we brought it up and I spent about a month with it and I ultimately killed the program." 00:32:40

Dr. Gustav Anderson

Hand surgeon in Sweden, recurring technical contributor to the Over The Horizon community. Provided clinically informed questions about hand kinematics, degrees of freedom, and the challenge of packing actuators and managing thermals in a robotic hand.

"Gustav Anderson is Dr. Gustav Anderson. He's a hand surgeon in Sweden. So great guy on Over The Horizon." 00:04:03

5. Operating Insights

Design Hardware Around the AI Model, Not the Other Way Around

Brett's most important organizational decision for Figure 4 was to reverse the traditional product development sequence — first define what the AI model needs, then design the hardware to serve that model's data requirements.

"We've designed figure four for Helix, Helix three was designed for data, and the whole robot was designed around data which is our like the largest limiting factor by far for figure to get into robots at mass scale." 00:18:27

More Diverse Data Beats More In-Domain Data for Improving Task Performance

When trying to improve a robot's performance on a specific task, adding unrelated-but-diverse training data can outperform adding more in-domain data. This is a non-obvious lever that can be pulled before requiring new task demonstrations.

"We re-exited the amount of data that was like nothing to do with the fridge — it was like opening cabinets and drawers and like things on the tabletop — and that same model I watched basically do an eliminated eval the next day basically do 90 in the fridge work. So we went from like 60 success rate with fridge only data." 00:13:58

Kill Local Maxima Early: Recognize When a Technical Approach Has a Hard Ceiling

Brett spent significant resources on a tendon-based hand design before personally shipping it, testing it for one month, and canceling the entire program. He frames this as correct decision-making, not waste — the lesson is to run fast experiments and be willing to terminate paths that are technically correct but structurally limited.

"Tendons is for sure the wrong way to go with hands. It's a local maximum and it is not the right engineering design long term for humanoids and I learned that firsthand... since then we've now worked on five or six new generations of hands." 00:32:40

Reproduce Customer Environments Exactly in Your Own Lab — No Simplifications

Figure deliberately replicates real customer site conditions in their internal testing environment, including the absence of guardrails on conveyors, package variety, and cube-shaped boxes. Designing for a sanitized internal environment creates capability gaps that only appear at deployment.

"We are trying to reproduce the exact use cases we see at real sites... one for one it doesn't help us designing systems here that don't look like the customer sites." 00:46:22

6. Overlooked Insights

Hark Is a 70-Person AI Lab That May Be Figure's Most Important Hidden Asset

Brett mentioned Hark almost in passing, describing it as a separate lab he started last summer with about 70 people. It is building real-time speech-to-speech models, next-generation AI architectures, and new hardware. Every Figure robot already runs the Hark voice model. This is not a small team working on a chatbot UI — it is a substantial AI research organization operating in parallel to the core robotics work, and almost no outside coverage exists on it. If Hark produces a breakthrough in real-time human-robot interaction or a novel model architecture, it could be a step-change capability that compounds the hardware advantage.

"I started a new lab called Hark last summer. We have a team about 70. We're working on like new kind of AI, next generation AI models and some hardware... every robot Figure now has the Hark voice model on it and you're able to stop any robot and talk to it here at the office." 00:49:49

Figure Has a Credible Path to Space Deployment and Is Already in Early Conversations

In two sentences buried at the end of the conversation, Brett confirmed that space deployment is part of Figure's master plan from founding, and that those conversations are happening now. Given the scale of demand NASA, SpaceX, and other space agencies will have for dexterous autonomous labor in environments inhospitable to humans — and given Figure's on-device inference capability that eliminates network dependency — this is not a distant fantasy. It is an addressable near-term market that almost no one is pricing into Figure's valuation discussions.

"On the space side — yeah we would love to be in space. It's part of our master plan from when we started and we're having some of those conversations now. Our robots will most certainly be in space and hopefully as soon as possible." 00:55:06