Smarty is still one of the hottest AI investment directions this year, but the industry's excitement is being pulled back to the ground by real problems: robotic hardware is advancing, and the “brain” that truly stabilizes a valuable job is not mature.
TechCrunch reported that this contrast had recently been evident. The Chinese robotics company, Woo Tree, once launched the market, but the stock price then fell sharply. Analysts believe that robots are improving their mobility very rapidly, but they are still far from a reliable, commercial task.
Industry heat continues to rise.
Actuate Congress last week brought together a lot of robotic AI developers. The host, Foxglove, states that since the launch of the conference in 2023, the size of the conference has tripled, reaching 1,500 participants this year.
But behind the heat, the core bottlenecks in the industry have not disappeared. Many practitioners have mentioned that robot AI currently lacks high-quality training data. It is difficult at this stage to deliver stable, replicable commercial performance, whether it is to be a universal robot or to conduct end-to-end training for a single mission.
Data and calculations remain bottlenecks
Antioch founder Harry Melsop describes the current period of "GPT-2" as being a smart one. His judgment is that this stage is less than a real breakthrough, more data and more calculus, especially for a real GPU.
As it stands, auto-driving companies are moving ahead. On the one hand, vehicles are naturally able to collect real road data on a continuous basis; on the other hand, the core task of automatic driving is more of a shield than a direct manipulation of complex objects. This has led to a more mature approach to autopilot in terms of data infrastructure, simulation systems and machine learning.
Autopilot to the machine. People
And that's why a lot of autopilots are moving their tools into robotics. Tesla's advanced Optimus, Wayve and Uber have also set up robotic laboratories to study human robotic forms.
Chief Executive Officer, Alex Kendall, believes that it is not yet appropriate to inscribe a single hardware platform prematurely because sensors and parts are still changing rapidly. He indicated that data infrastructure, simulations and ML Ops tools would have a high commonality, but that different robotic forms would still require their own post-training and world model adaptation.
However, co-design of software and hardware was also advocated. Chief Executive Officer Genesis AI Théophile Gervet stated that the industry was too early, that the “robots brain” strategy alone was not always sufficient, and that there was still considerable room for joint hardware and AI design.
Universal human form still in the lab.
Industry has become fragmented on the commercial path. Mission-oriented robots are starting to enter real scenes, such as Gritt's involvement in solar power, Agility's deployment of robots to industrial scenes and Bedrock's automating excavators. In contrast, most generic human-shaped robots remain at the laboratory level.
Gervet, to be blunt, does not need a universal robot with an 80% success rate. Without clear scenarios, it is difficult for products to create real value. But if only a very narrow vertical task is done, it may lose its advantage when the next generation of stronger models emerges.
In the context of data management, Foxglove also released a new product this week that allows engineers to retrieve visual and laser radar data in natural languages for faster assessment, detoxification and simulation based on the British Columbia Cosmos Open World Model.
As to when intelligence comes to a turning point like ChatGPT, there is no uniform answer within the industry. It is expected that low-cost, highly autonomous products will be available at the consumption end first, and it is argued that robots will not have flash time like software, but will be able to enter home and industry more slowly, as in the early days of personal computers.