Brett Adcock, CEO of the robotics company Figure, called Helix 2.5, the neural network that controls the humanoid robot Figure 03, "the Holy Grail of robotics." This term refers to a highly sought-after achievement in the field, and on September 17, 2026, Figure demonstrated the system in a series of tests. During the demonstration, Figure 03 was tasked with performing everyday household chores such as tidying up a living room, folding towels, and making beds in 30 different apartments across the San Francisco Bay area. Importantly, the robot had no prior knowledge of the rooms or the objects it encountered, making the tasks more challenging.
To prepare for these tasks, Helix 2.5 was pre-trained using Index, a large database of filmed human behaviors that Figure made public in late August 2026. From this data, the company selected three specific tasks for testing: tidying up, folding towels, and making a bed. The term "zero-shot" refers to the robot’s ability to perform these tasks in completely new environments without any prior exposure to the specific objects or layouts. This is a significant challenge in robotics, as it requires the system to rely on general knowledge rather than specific training for each environment.
According to Adcock, the robot successfully completed about 56 percent of the trials, a marked improvement over a model with no prior training, which succeeded only 9 percent of the time. The company also recorded nearly four hours of footage showing the robot performing tasks across the 30 houses. Despite this progress, Figure acknowledges that the robot still falls short of the capabilities of a human butler. The company did not count partially completed tasks, such as a bed that was only half-made or towels that were not fully folded, emphasizing that the robot must complete tasks entirely to be considered successful.
Figure describes Helix 2.5 as the most advanced neural network it has developed to date. The company has attracted significant investment, having raised over one billion dollars in funding in September 2025, which valued the company at 39 billion dollars. While Figure does not claim to have fully solved the challenges of domestic robotics, it highlights that Helix 2.5 represents a major step forward. The system shows that a humanoid robot can apply skills learned in one setting to completely new environments without needing to relearn everything from scratch, marking an important milestone in the field of robotics.
Figure Robotics Demonstrates Humanoid Robot's Ability to Perform Tasks in Unfamiliar Environments
AI-rewritten from original reportingHow it works
roboticsaineural-networkfigure-03automationzero-shot-learning
Original sources:
- 🇫🇷Numerama



