Before AI robots enter your home, they're learning to work in warehouses
Digit walks into an Amazon warehouse on two legs. The bipedal robot picks up containers and carries them to conveyor belts, working three shifts a day with occasional breaks to recharge. It moves through the same stairs and hallways its human colleagues do, without requiring a million-dollar facility retrofit.
But there's a tradeoff to this freedom of movement. “They are balancing all the time — which means they could fall,” Jonathan Hurst, co-founder of Agility Robotics, which designed Digit, told the Infinite Loop by Nebius. To prevent a painful landing on a human worker, a barrier separates the robot's work cell from its teammates. That barrier is just the start of a multi-layered safety net woven into Digit's architecture — and into every AI robot now entering workplaces alongside humans.
Layer 1: Built for human spaces
Agility Robotics, an Oregon-based company, built Digit to work in spaces built for people. The bipedal vaguely resembles its human colleagues, but is a long way from a replicant. Hurst, who’s also Agility’s chief robot officer, describes the design as “human-centric.” Proportions and gait are tuned for balance, with a lower center of gravity and wider stability margins than a human’s. This supports dynamic stability: the ability to maintain balance while moving through complex environments.
Other robots take a different approach to physical safety. Locus Origin and Kachaka don’t balance on dynamically stable limbs; they move on wheels. This form lacks the versatility of Digit but enables work in the same spaces as humans.
Kachaka, built by Japan's Preferred Robotics, sports a rounded, compact design that can navigate tight spaces and carry configurable payloads. It delivers mail and small packages through KDDI's Tokyo headquarters, carrying loads that can be reconfigured for different tasks.
The Origin, developed by the Massachusetts-based Locus Robotics, functions like a mobile shelving unit, rolling through DHL distribution hubs to collect items from warehouse workers. Each Origin carries onboard AI that perceives obstacles, avoids collisions and plans its path. When multiple Origins work together, they share a collective intelligence through Locus Robotics' LocusONE platform, which coordinates the fleet like an air traffic controller for robots.
These are the environments where robots learn to coexist with people — the proving ground for everything that comes after.
Layer 2: Learning to see people
The physical body is only the first layer of safety. For robots to work alongside humans, they need more than stable forms — they need situational awareness. This is where the second layer comes in: the brain.
“The most fundamental challenge is detecting obstacles, especially people, in real time,” said Kane Edwards, business development manager at Locus Robotics.
The Origin combines LiDAR sensors to detect people and objects with depth-sensing 3D cameras to spot hazards that standard sensors miss, from dropped items to shifting floor levels. Even then, gaps remain that AI must bridge.
A common blind spot: raised forklift forks and elevated operator compartments. Standard sensors struggle to detect obstacles at varying heights. The Origin's AI-driven object recognition fills these blind spots, Edwards said.
When multiple Origins work together in a warehouse, they operate as a coordinated fleet with a collective mind. The robots share their intended routes through LocusONE, which acts like an air traffic controller — monitoring traffic, predicting congestion and adjusting paths before dangers form. This coordination uses multi-agent reinforcement learning.
“Without predictive path planning, where robots essentially share their intended routes and adjust proactively, they end up constantly stopping or rerouting on the fly,” Edwards said. “This creates unpredictable movements that can unsettle nearby workers.”
Digit's brain works differently. Using NVIDIA’s Isaac Sim application, Agility trains a whole-body control foundation model on decades of simulated time in just days. It’s then deployed "zero-shot" to Digit, creating an "always on" safety layer that instinctively manages disturbances like bumps and pushes.
Kachaka takes a more conservative approach. Its AI plays a supporting role, handling perception but not movement decisions. Tomo Toru Isobe, CEO of Preferred Robotics, said this separation keeps high-stakes movement under deterministic control rather than AI decision-making.
Kachaka analyzes camera feeds pixel-by-pixel using deep learning, identifying walkable areas and obstacles that LiDAR sensors often miss. Specialized SLAM (Simultaneous Localization and Mapping) adds spatial awareness, while a fleet management system provides coordination.
Every sensor, every algorithm, every edge case encountered in a Tokyo office building is data that makes the next generation of robots safer in your living room.
Layer 3: Rules before rooms
Physical design and AI perception alone aren't enough. Digit, Origin and Kachaka can only work where formal boundaries permit them — the safety net’s third layer.
Before Kachaka could enter KDDI's offices, it underwent Failure Modes and Effects Analysis (FMEA) — a process that analyzes every potential failure and its consequences. This rigid engineering framework, required by Preferred Robotics, ensures robots are bound by formal safety processes before they’re shipped.
Regulations and certifications are also stitched into the safety net “from an early stage of development,” Isobe said.
Locus Robotics must also comply with various standards, yet they’re not merely burdens. Paradoxically, these boundaries can ease access to international markets. Take CE certification, a mark of EU safety compliance that the Locus Vector robot achieved last year. “It's compulsory for products entering the European market, so achieving it was essential for serving our international customers,” Edwards said.
Digit faces a unique challenge: no safety standard existed for bipedal, dynamically stable robots. Agility Robotics is leading the development of ISO 25785-1 — the first international safety standard for robots like Digit — while building a certification scheme for insurers. The same standards that govern Digit in an Amazon warehouse will eventually govern the robot that folds your laundry. “We need a way for insurers to understand the risk they are underwriting,” Hurst said. “The best way to do that is with an industry-wide standard everybody agrees on.”
Even with these safety layers, robots like Digit still can't work in every environment humans do. Humanoids in households, for instance, are over 10 years away, Hurst said.
Safe deployment alongside humans remains the biggest barrier. Homes are immensely complex, variable environments with unpredictable children, pets, narrow hallways, and wet floors. No company wants its robot to fall on a child.
For now, warehouses, construction sites, and offices are where that problem gets solved, one edge case at a time.
“At some point, you can get them in the home,” Hurst said. “But it's going to be after all of these industries.”
The three-shift days in Amazon warehouses, the mail runs through KDDI's Tokyo offices, the path-planning algorithms navigating DHL hubs — all of it is rehearsal. The home is the final exam.
This story was produced by the Infinite Loop by Nebius and reviewed and distributed by Stacker.
