Written by: Techub News Compilation

Introduction

In June 2025, Brett Adcock, the founder and CEO of humanoid robotics company Figure, was a guest on the tech podcast “Brighter with Herbert,” engaging in an in-depth conversation lasting an hour. The crux of this discussion was the first public presentation of a complete, unedited 60-minute video showcasing the work of the Figure 02 humanoid robot. In the video, the robot continuously sorted small package logistics, drawing widespread attention within the industry for its smoothness and stability. Brett Adcock not only interpreted the technological breakthroughs behind the video but also delved deeper into the real challenges faced by humanoid robot commercialization, the company’s strategy, and prospects for future applications. Given the current trend where demonstration videos in the humanoid robotics field predominantly consist of short clips, this long, unedited live demonstration marks a significant milestone.

Summary

• The Figure 02 robot is powered by the single neural network Helix S1 and achieves a high success rate in complex small package sorting tasks with just about 60 hours of training data.
• The key to success lies in fundamental improvements to the Helix model, including the introduction of stereo vision, force feedback perception, and time memory (spatial-temporal memory) capabilities.
• The company’s goal is for humanoid robots to perform “tedious, dirty, and dangerous” repetitive labor, with a core advantage in versatility, allowing adaptation to the human-designed world and continuous evolution through a data flywheel.
• The next-generation robot, Figure 03, is already in internal testing phases, with significant improvements in speed, vision, and hand dexterity, and the company claims there are currently no supply chain bottlenecks.
• Brett Adcock believes that the collaborative design of hardware and AI, along with scalable data collection and manufacturing capabilities, is key to winning the humanoid robot race.

From “Show Off” to “Practical Action”: An Hour of Silent Declaration

Unlike the carefully edited, short videos showcasing parkour or dance released by many robotics companies, Figure chose to present an unedited, single-angle raw video lasting 60 minutes. The video content is even somewhat “tedious”: the Figure 02 robot stands by a conveyor belt, continuously identifying and grasping small packages (including plastic bags and boxes) from a chaotic pile of packages, flattening them, flipping them over so that the barcode is facing down towards the scanner, and then placing them in designated positions. The entire process is repetitive.

It is precisely this “tediousness” that becomes the most powerful technological declaration. Brett Adcock pointed out that small package logistics sorting has always been the “Achilles' heel” of general-purpose robotics. Plastic bags deform when grasped, and package states never repeat, making it impossible to address with traditional programming and heuristic algorithms. This is a problem “perfectly suited for neural networks.” The video demonstrates that Figure 02 can perform this complex task steadily and continually with a speed close to human levels (approximately 4-5 seconds per item at the time, with the latest data reaching 3.54 seconds per item) and an over 95% success rate.

Brett Adcock emphasized that Figure's mission is to “sell human work,” meaning allowing robots to replace humans in engaging in practical labor in logistics, manufacturing, construction, and eventually entering households. As such, their priorities are output, reliability, and return on investment rather than demonstrates functionalities. The robustness and durability portrayed in this one-hour video are critical indicators for assessing whether robots can be deployed in practical scenarios.

The Evolution of the Helix Brain: Vision, Touch, and Memory

The core of this performance breakthrough is Figure's independently developed AI system, Helix. Brett Adcock detailed several fundamental improvements made to the Helix S1 model for this task:

1. Stereo vision and perceptual memory: The model employs stereo cameras for input images and introduces a “time memory” capability. The robot can remember information from camera frames from the last few seconds. This is crucial in a dynamically changing environment. For instance, when the robot turns its head to perform actions, it may temporarily lose sight of packages on the left, but thanks to memory, it knows there are still packages to grasp. Brett Adcock even observed behaviors similar to humans, such as the robot stretching its arm behind it to grasp packages that had moved out of view.

2. Force feedback integration: Incorporating force sensor data into the model's state input allows the robot to perceive grasping strength, enabling more dexterous handling of objects like deformable plastic bags.

3. Expanded model scale: Helix S1 itself is a larger model in terms of parameter size, and this scale increase directly results in performance enhancements.

These architectural improvements, combined with approximately 60 hours of targeted training data, have collectively contributed to significant increases in robot speed and success rate. Brett Adcock highlighted that these neural network-level improvements are general and will benefit all of the company's robots and application scenarios, creating a strong technological synergy.

Outpacing Humans: The Race for Speed, Success Rate, and Versatility

During the conversation, the host and guests delved into the advantages and disadvantages of humanoid robots compared to traditional automation solutions (such as specialized robotic arms). Brett Adcock and the guests presented several core arguments:

Market reality: Currently, millions of workers globally are engaged in similar repetitive sorting tasks rather than fully automated robotic arms. Many small and medium-sized enterprises do not have sufficient capital to invest in highly customized, specialized automation production lines. Humanoid robots offer a versatile, redeployable solution.

Versatility advantage: Specialized robotic arms can only perform singular tasks. The humanoid form factor is inherently adaptable to a human-designed world (tools, stairs, doorways, etc.). By employing a general “body” (humanoid platform) and “brain” (like Helix), they can theoretically learn to execute countless tasks. Skills learned in a logistics context, such as grasping and recognition, can be transferred to other scenarios like manufacturing and household services, achieving “transfer learning.” Brett Adcock believes humanoid robots are the “ultimate deployment vehicle for general artificial intelligence.”

Key performance indicators: Beyond the speed and success rate of singular tasks, true commercialization must assess other KPIs: uptime/failure rate: Can the robot work continuously for 8 hours, days, or even months without human intervention or maintenance? Durability: The robot’s lifespan and maintenance costs. Human replacement ratio: How many human employees can one robot replace? Considering that human employees need breaks and have high turnover rates (Brett Adcock mentioned such positions can see an annual turnover rate of up to 100%), robots, even if their speed is temporarily on par, possess the potential for significant overall efficiency gains through 7x24 hour continuous operation.

Brett Adcock disclosed that during the demonstration, the robot achieved a processing speed of 3.54 seconds per package, reaching the average level of “excellent human employees.” Since the design peak speed of Figure 02's actuators far exceeds the speed currently utilized, he believes that through software optimization, the speed could drop below 3 seconds within the next few months, potentially breaking 2 seconds by the end of the year, thus greatly surpassing human speed limits.

Figure 03 and the Path to Scaling: Hardware, Data, and Manufacturing

Despite the impressive performance of Figure 02, Figure’s development has not paused. Brett Adcock excitedly revealed that the next-generation robot, Figure 03, is already in internal testing, and progress has far exceeded expectations.

Hardware upgrades: Figure 03 has been optimized in multiple aspects: the speed of finger movements is twice as fast as Figure 02 to meet high-frequency grasping demands; a custom-designed camera has significantly enhanced visual perception capabilities; and significant improvements have been made to the hands. The overall goal is to get the robot’s range of motion, speed, and payload (torque) closer to that of an average human to cover a broader array of physical labor scenarios.

Rapid iteration: Thanks to team experience accumulation and software stack (like Helix) reuse, Figure 03 took less than a week to go from assembly to achieving bipedal walking, while Figure 02 took three months. Currently, Figure 03 can complete some basic fully autonomous task processes.

Scaling challenges: Brett Adcock believes the current main bottlenecks lie in AI learning and scalable manufacturing. Regarding AI, the core issue is how to efficiently and widely collect training data. The advantage of humanoid robots is that their form is consistent with that of humans, allowing for efficient learning from human demonstrations. In terms of manufacturing, Figure has established its own production facility, “Bot Q,” and has begun producing Figure 03. Brett Adcock confidently stated that the company currently has “no supply chain issues” with Figure 03, having considered high-rate production from the design stage, for instance, its injection-molded head component line can reach a weekly capacity of 50,000 units.

He described a powerful flywheel effect: deploying more robots to work in real scenarios will allow them to collect massive data (including successful and failed cases), which will be fed back to Helix for retraining, making all robots smarter. Smarter robots can handle more tasks, create broader demand, and further drive larger-scale manufacturing and lower costs. “The smartest robots will also be the cheapest robots,” Brett Adcock summarized, while the company capable of achieving rapid deployment of 100,000 or even a million units will gain a significant first-mover advantage.

Future Outlook: From Factories to Homes

When asked when robots could enter homes to complete tasks such as “assembling IKEA furniture” or everyday washing dishes and laundry, Brett Adcock provided an optimistic forecast. He believes that as long as hardware capabilities (such as range of motion and strength) are up to standard, these tasks can be learned through data training. He anticipates that robots capable of performing daily household chores could appear “within the next two to three years,” while more complex tasks like reading instructions and assembling furniture will also be realized “within this decade” (i.e., in the coming years).

Finally, Brett Adcock refuted the viewpoint that “AI is a commodity that can be easily transplanted to any hardware.” He emphasized that humanoid robot hardware itself is an extremely high-dimensional, complex system that is challenging to control. High-quality AI needs to be co-designed with dedicated, high-performance hardware to maximize efficiency. He believes that globally there may ultimately be only a few companies (a few each from China and the U.S.) that can truly address the comprehensive challenges of humanoid robot hardware, AI, and scaling deployment.

This dialogue and the one-hour demonstration video together conveyed a clear message to the world: humanoid robot technology is steadily moving from a “show-off” phase in laboratories to a “practical action” phase focused on solving real economic problems and creating commercial value. The challenges ahead remain substantial, but the path is becoming increasingly clear.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。