A Robot That Actually Cleans Up Trash with Just the Command "Clean Up the Trash"—Agility's CEO Reveals the Division of Labor Strategy Between "Brain" and "Body"

With just the instruction "Clean up the trash," Digit correctly identified bubble wrap as non-recyclable and sorted various pieces of trash scattered on the floor into bins. This demonstration, revealed in an interview with TechCrunch in early July by Agility Robotics CEO Peggy Johnson, is quietly exciting for software engineers.

Agility recently announced its IPO plan through a SPAC merger (valued at approximately $2.5 billion, with over $620 million expected to be raised), but this time, we want to focus not on the glamorous fundraising, but on the design philosophy of the technology architecture discussed in the interview.

A Smart Principle of "LLM Independence"

First, what's interesting is Agility's clear stance on LLM (Large-Scale Language Models): "LLM-agnostic" (not dependent on LLM). They use multiple models, such as Claude and Gemini, depending on the situation.

Johnson explains that the role of LLM is limited to the "semantic layer"—that is, the part that translates high-level human instructions ("clean up the trash") into a concrete sequence of actions that the robot can execute. This is a very logical design. Being tied to one company's LLM means bearing the entire risk of that model's performance, price, and availability. Conversely, by focusing on the abstract task of "semantic understanding," they can ensure the flexibility to switch to the latest high-performance model as it becomes available.

The "physical layer," which LLMs struggle with, is the heart of Agility

This was the point that resonated with me the most in this interview. Johnson emphatically states, "LLMs could use the entire internet as training data. But for humanoid physical AI (AI with physicality), such data doesn't exist yet."

In other words, the idea is that physical control—such as walking while maintaining balance, grasping objects, and carrying heavy loads—cannot be acquired through text or image training, but can only be honed through years of accumulated real-world operational data. Agility, a company spun off from Oregon State University in 2015, has over 10 years of experience operating robots in actual warehouses and factories. Johnson confidently states, "We may have one of the largest data lakes for real-world robotics data in actual operating environments."

This is something I always think when watching the flashy demo videos from Chinese companies: Visual smoothness and reliability in actual operation are two different things. Agility's approach, which differentiates itself from flashiness by focusing on "physical layer data accumulated over 10 years," feels like a grounded strategy.

Digit's Hardware: Surprisingly "Pragmatic" Design

I'd also like to introduce the specifications of the Digit robot itself. Height approximately 175cm, weight approximately 73kg. The most distinctive feature is its reverse-jointed knees (so-called "bird legs"), designed to allow it to reach from floor level to overhead shelves without hitting warehouse racking. According to Johnson, this wasn't simply biomimetic, but rather a result of working backward from the desired function.

The hands also have a simple configuration of "two thumbs and two fingers," optimized to securely grasp heavy plastic containers even if their contents shift during transport. This design philosophy, which focuses on minimum degrees of freedom specific to the task rather than directly mimicking human hands, resonates strongly with me as a software engineer. Excessive versatility often only increases complexity and cost, without directly translating into practical operational benefits.

The "Unassuming but Inescapable" Hurdle of Safety Certification

Another point Johnson emphasized was safety. While competitors showcase robots through lab demonstrations and choreographed videos, Agility has been forced to meet actual industrial safety certification requirements in order to operate within customer facilities. Her statement, "You can't make a robot safe after you've built it. That would require a complete redesign. Everything—the electrical systems, components, and even the supporting software—needs to be safety certified," resonates with the NVIDIA Halos story I previously discussed in this column, illustrating a common industry challenge.

In this context, I'd like to mention the case of Figure AI's former product safety manager, who sued after raising concerns that the company's robots "possessed enough power to potentially fracture a human skull" (Figure AI contests the claims). Regardless of the veracity of the allegations, the very fact that such lawsuits are occurring demonstrates that humanoid safety design is still a developing area.

A Realistic Approach to "Household Robots"

Finally, another aspect of Ms. Johnson's remarks that I found appealing was her stance on household humanoids. He explicitly stated that it would take "more than 10 years," explaining that "while warehouses and factories are complex, the aisles are fixed and the equipment layout is predictable. Homes, on the other hand, have dogs, babies, guests, and things end up in unexpected places—it's a completely different story."

His comparison to self-driving cars, noting that "roads still have a certain order," is quite grounded, especially considering the often exaggerated predictions in this industry. Currently, Agility is focusing on the more than one million labor shortages in the US, created by retiring workers and young people avoiding physically demanding jobs. Throughout the interview, his focus on the concrete challenge of labor shortages, rather than the flashy "dream of household robots," was consistent.

Summary

While the humanoid industry often attracts attention for its glamorous fundraising competition (such as Figure AI's $39 billion valuation and Apptronik's $5.5 billion valuation), this interview teaches us the healthiness of a design philosophy that separates the brain (LLM) from the body (physical control). By leaving the evolution of the LLM to other companies and focusing on the physical layer data and safety certifications accumulated over 10 years—assets that are not easily replicated—Agility's pragmatic approach offers valuable insights for companies considering introducing humanoids into factories and warehouses.