[Physical AI Verification Case] Verification of Practical Feasibility of Quadruped Robots Using AI

Athena Technologies Co., Ltd. (Headquarters: Bunkyo-ku, Tokyo; Representative Director: Takeshi Abe; hereinafter 'Athena') and Tokyo Electric Power Company Holdings, Inc. (Headquarters: Chiyoda-ku, Tokyo; Representative Executive Officer and President: Tomoaki Kobayakawa; hereinafter 'TEPCO HD') have conducted a verification of an 'AI Agent Control Pipeline' that enables users to operate quadruped robots using only voice commands.

In this verification, a system was built to recognize and interpret voice commands, convert them into operational commands for the quadruped robot, and execute them on the actual machine. This verified the potential for a more natural robot operation interface based on voice, in addition to standard robot operations.

■ Two Key Points Verified for Utilizing Quadruped Robots

Quadruped robots can be remotely operated via tablet applications and also allow for robot control and status information acquisition from applications using an SDK. Based on these existing functions, this verification developed and tested the following two points to achieve a more user-friendly operation interface.

Connecting Natural Language Instructions and Robot Operations

The possibility of application is being explored by enabling language-based operations such as 'Patrol' and 'Check for anomalies.' For operation by an AI agent, voice instructions must be translated into executable operation commands. In this verification, a pipeline was built to interpret the content of instructions obtained through voice recognition, refer to predefined quadruped robot operation commands, and generate code.

Execution Control Aiming for Improved Stability and Safety

In robot control, it is important to appropriately control the execution scope and operation procedures to avoid unintended actions. In this verification, instead of having the AI generate code without constraints, the available operations were limited, and the configuration connected voice input to the actual execution of the quadruped robot, aiming to ensure operational stability and safety.

Athena has been verifying the potential of a new interface that uses AI for voice recognition and intent interpretation, connecting natural language instructions to the operation of quadruped robots.

■ Confirmed Execution of Predefined Actions with Voice-Controlled AI Agent for Quadruped Robots

This verification mainly involved the following three initiatives.

Initiatives in this verification

Verification Results

End-to-end pipeline streamlining from voice recognition to action script generation

Confirmed execution of predefined actions

Clarification of ambiguous voice instructions through a clarification function

Improved code generation success rate

Implementation of an AI agent for code review

Minimized risk of malfunction

Construction of a Voice Control Pipeline by AI Agent

A pipeline was built that performs voice recognition, intent interpretation, task decomposition, tool selection, and code generation. Evaluation was conducted on predefined action scenarios, and execution was confirmed for each target action. However, challenges remain in operational stability at this stage, and it was confirmed that behavior can vary depending on the usage environment and conditions.

Clarifying Ambiguous Instructions by 'Asking Back'

For abstract voice instructions unique to the field, such as 'Patrol,' a 'clarification function' was implemented where the AI automatically identifies missing information and asks the user for confirmation. Compared to the code generation success rate without the clarification function, the success rate with the clarification function was significantly improved. The more ambiguous the instruction, the greater the effect of the clarification function, enabling the robot to more accurately understand the user's intent.

Ensuring Operational Stability by Predefining Code

Instead of having the AI write code from scratch, several actions (walk, take a picture, grasp, etc.) were predefined as Tools. By referencing these tools, code generation is possible by combining only the procedures within them. This design minimized the risk of malfunction.

■ Preventing Robot Malfunctions with Multi-Stage Safety Design

In this system, the system was designed so that the AI accurately understands voice language instructions and then selects the appropriate action from a set of predetermined actions.

This aimed to reduce the risk of the AI causing unexpected actions and ensure operational stability.

■ Potential for Future Technological Expansion

Through this verification, the effectiveness of intuitive voice-based operation was confirmed. Based on these results, the goal is to evolve quadruped robots from mere command-executing robots into 'AI agents' that can make their own decisions and provide explanations. The following functional developments are being considered to aim for further expansion.

Improved Operational Efficiency through Increased Autonomy

Improved applicability to more complex tasks by enabling the robot to judge the situation and modify its actions, such as automatic re-photography upon failure or autonomous charging based on battery level.

Re-planning According to the External Environment

Development of a function where the AI agent recognizes the external environment using image data captured by the quadruped robot and re-plans accordingly.

Improving Instructions by Referencing Reference Materials

Verification of a function that enables control instructions aligned with the task by performing knowledge retrieval using RAG before converting voice instructions into robot control code.

From a 'robot that executes commands' to a 'robot that thinks, judges, explains, and can operate autonomously even in closed environments.' Based on the results of this verification, we will continue to aim for a future where Physical AI becomes a partner in the field.

■ Representative Comment

Takeshi Abe, Representative Director and CEO

'To improve safety and efficiency in the field, we need an interface that breaks down barriers of specialized knowledge and allows anyone to operate intuitively. Through this project, we were able to confirm that the AI agent understands the 'voice' of field workers and executes predefined actions. I am confident that the technology to 'control robots with voice,' once a concept in science fiction, is already reaching a practical level. Going forward, based on the knowledge gained from this verification, we will accelerate verification and development aimed at realizing 'Physical AI' that functions in actual field settings, not just in laboratory-level verifications.'

■ Company Overview

[About Athena Technologies Co., Ltd.]

As a 'Matsuo Lab-born startup (*1),' we combine cutting-edge academic technology with social implementation capabilities, offering secure AI development and governance support.

Company Name: Athena Technologies Co., Ltd.

Established: December 2023

Representative Director: Takeshi Abe

Location: HONGOEGG, 6-25-14 Hongo, Bunkyo-ku, Tokyo (Tokyo Office); Hyakumanben Building 3F, Tanaka Monzen-cho, Sakyo-ku, Kyoto City, Kyoto (Kyoto Office)

URL: https://athenatech.jp/

(*1) 'Matsuo Lab-born startup' refers to a select group of startups founded by alumni of the Matsuo Laboratory or established with support from the Matsuo Laboratory, whose technology and business capabilities are recognized for growth potential, who share the philosophy of the Matsuo Laboratory, and who work together to foster the next generation. (Registered Trademark No. 6667237)