TorqueAGI 稱機械人學習新任務所需數據最多可減少 1,000 倍

One of the key challenges in bringing physical artificial intelligence (AI) to manufacturing is securing training data for robots. Yet collecting data for robotic AI is far from easy. Data with high training value, particularly near misses, production-line stoppages and abnormal situations, is difficult to collect deliberately because of safety and cost concerns. The shortage of high-quality data is widely regarded as one of the biggest obstacles to innovation in physical AI.

To overcome this data shortage, synthetic data generated through digital twins has emerged as an alternative. A further development is self-learning, in which AI trains itself using curated synthetic data while continuously reducing the gap between virtual and real-world environments in real time.

Whether through teleoperation to collect real-world data or synthetic data generated in virtual environments, these approaches share a common assumption: robotic AI needs sufficient data to learn effectively. U.S. physical AI startup TorqueAGI is challenging that premise. Its technology seeks to address the problem by rethinking why robotic AI needs so much training data and equipping robots with the ability to reason about the physical world.

In conventional systems, when a robot encounters a new object or working environment, it typically needs to collect and label data specific to the new conditions before being retrained. When a new product is introduced to a production line, for example, the robot must learn its position, shape and appropriate grasping method again.

TorqueAGI incorporates an object's shape, weight, geometric relationships, physical forces and operating constraints into the model's reasoning process. Rather than memorizing images of specific objects, the model is designed to reason about how objects move and respond to forces. This approach can enable robots to learn new tasks using up to 1,000 times less data than conventional methods, according to Ashutosh Saxena, chief executive officer (CEO) of TorqueAGI.

TorqueAGI calls its core technology a "World-Reasoning Model." Saxena said the company's model incorporates physics and geometry into its reasoning process. The approach can be understood as combining a world model with physical reasoning and robotic control.

Saxena, who also founded TorqueAGI, previously conducted research at Stanford University on Make3D, a system that reconstructs three-dimensional structures from a single image, as well as deep learning-based robotic grasping. At Cornell University, he subsequently developed RoboBrain, a system that enables different robots to share knowledge acquired through learning. These research projects share a common objective: enabling robots to understand the shapes and spatial relationships of objects and transfer acquired knowledge to other tasks and robots.

TorqueAGI's World-Reasoning Model goes beyond training robotic AI to recognize visual patterns and reproduce previously learned tasks. It is designed to understand object shapes and weights, machine structures and work procedures within spatial and temporal contexts, using physical principles such as gravity to determine what action to take next.

For example, if the size of a gripper changes or the position of a conveyor shifts, the robot does not need to relearn its entire operation from scratch. Instead, it calculates a new grasping method and movement trajectory based on the changed conditions. The approach aims to reduce dependence on training data by giving the model a foundational understanding of the laws of physics.

According to TorqueAGI, its World-Reasoning Model enables a world model to understand and reason about physics and geometry.

Using TorqueAGI's World-Reasoning Model, tasks that would traditionally require tens of thousands of demonstrations can be learned with just dozens of examples, alongside existing customer data such as equipment records and operating procedures, Saxena said.

In one case, a customer in the heavy equipment sector reportedly achieved comparable performance on a task using approximately 6,000 data points with TorqueAGI's model, compared with about 7 million event records required to train the previous model. Saxena also highlighted the ability to reduce data collection for new tasks to just a few days and the model's explainability, which allows users to understand why it selected a particular action.

This has important implications for high-mix, low-volume (HMLV) manufacturing and customized production. As manufacturers increasingly move toward HMLV and customized production, the limitations of industrial robots programmed to repeat fixed movements are becoming more apparent. The biggest obstacles to autonomous manufacturing and robot transformation (RX) are often not the cost of the robots themselves, but the expenses associated with reprogramming, data collection and validation every time production tasks change.

If robots can adapt quickly to new components and tasks while manufacturers retain their existing equipment, automation can be applied to a wider range of operations and production-line changeover costs can be reduced. Saxena said South Korea's manufacturing industry, which already has extensive automation infrastructure, represents a promising market for TorqueAGI's technology, as physical reasoning could make existing robots and production lines more flexible.

Saxena also said the company's World-Reasoning Model could increase packing density by understanding geometric characteristics and reasoning about spatial relationships. By considering the weight, shape and fragility of individual items, the model determines where and in what order products should be placed. This spatial reasoning allows it to use available space more efficiently and increase packing density.

TorqueAGI sees opportunities for its solutions in three areas of South Korea's electronics and components manufacturing industries: inspection and quality control (QC), material handling and logistics, and assembly. The respective solutions are TorqueFocus, TorqueFlow and TorqueBuild.

TorqueFocus can perform complex assembly verification, defect detection, wire harness measurement, equipment monitoring and safety inspections. TorqueFlow automates tasks such as component sorting, kitting and packaging inside factories. TorqueBuild is designed for component insertion, fastening and multi-stage assembly operations. The TorqueBuild solution is currently under development.

TorqueAGI develops models jointly with customers and then distills them for deployment on NVIDIA edge computers installed directly on machines. The company uses platforms including NVIDIA Orin and NVIDIA Thor. Its systems are designed to maintain machine perception and safety-related decision-making even when network connectivity is lost during operations.

Saxena emphasized that customers can operate independent models in their own edge environments, ensuring that one company's data is not used to train robots belonging to another company. He said this architecture could be particularly suitable for South Korean advanced component manufacturers, including semiconductor companies, where data security is critical. While cloud infrastructure is used for model development and updates, each customer operates an independent model, enabling on-site inference and operational decision-making without transmitting manufacturing data externally.

However, Saxena acknowledged the challenges involved in actual deployment, saying that the model cannot do everything on its own. "The model itself changes very little even when the machine changes, but most of the work takes place on the machine side," he said. He added that integration takes time, including connecting the model to a manufacturer's control interface, calibrating sensors and ensuring that the model complies with the machine's safety systems and operating limits.

TorqueAGI has secured proof-of-concept results in logistics, warehousing and heavy equipment applications. In warehouse packaging operations, the company said it processed approximately 700,000 orders, considering factors such as product weight and shape, and increased packing density from 73% to 93%. However, the company has yet to secure proof-of-concept results in electronics and semiconductor manufacturing.

TorqueAGI began engaging with the European market in September and is currently building business relationships in Asia as well. In South Korea, the company is pursuing market entry through robot manufacturers and system integrators (SIs). Its strategy is to conduct joint proof-of-concept projects on specific tasks at actual production lines, using performance indicators agreed upon in advance with customers.