Figure Helix 2.5 is taking the Figure 03 humanoid robot into homes it has never seen before. Figure tested the system across 30 unfamiliar homes without collecting training data inside those evaluation environments.
Helix 2.5 demonstrates Figure 03 performing demanding household tasks including making beds, folding towels, and tidying living rooms. The robot must combine artificial intelligence, vision, locomotion, manipulation, coordination, and full body movement to complete these jobs in unfamiliar surroundings.
One of the biggest developments behind Helix 2.5 is Figure's Index dataset of human physical behavior. In Figure's controlled testing, a policy without Index pretraining achieved a 9% zero shot success rate, while the Index pretrained Helix 2.5 policy reached 56%, representing an improvement of more than 6 times.
The robot can also recover from problems during longer tasks by repositioning its body, changing its stance, stepping backward, and moving around furniture to continue working.
This does not mean humanoid robots have completely mastered household work. Instead, the demonstration shows progress toward robots that can transfer previously learned physical skills into new environments without requiring training inside every individual home.
Could you see yourself having a Figure humanoid robot handling chores around your house? Let us know in the comments, and remember to like the video and subscribe to DPCcars for more humanoid robotics, artificial intelligence, automotive technology, and future technology news.
Figure is pushing humanoid robotics deeper into the home with Helix 2.5,
its latest artificial intelligence system designed to help Figure 03 handle real household work in
places the robot has never seen before. Instead of preparing the robot for one carefully controlled
room, Figure took it into 30 unfamiliar homes and asked it to get to work. The challenge sounds
simple until you consider how different every home can be. Beds have different heights,
furniture changes position, rooms can be tight or cluttered, and everyday objects come in countless
shapes and sizes. Humans naturally adjust to these differences. Robots usually have a much harder
time. Helix 2.5 was designed to close that gap. Figure tested the system with three demanding
household behaviors, making beds, folding towels, and tidying living rooms. These jobs require much
more than recognizing an object and picking it up. Figure 03 has to move through the room,
position its entire body, use both hands, understand what it is seeing, manipulate soft
materials, and continue working through a sequence of actions. Most importantly, Figure says no
training data was collected inside any of the 30 evaluation homes. The robot also encountered
objects that were not part of the task-specific training data. It had to rely on what it had
already learned and apply that knowledge to the new environment. That is the key idea behind the
zero-shot demonstration. Helix 2.5 was trained for these types of behaviors elsewhere, but it was not
retrained for each new house. Figure says a major reason for the improvement is Index, its large
data set of human physical behavior. Helix 2.5 was pre-trained using this broad collection of
human experience before being adapted for the individual household tasks. The difference was
substantial. In Figure's controlled comparison, a policy trained without Index pre-training
achieved a 9% success rate during zero-shot testing. With Index pre-training, the success
rate increased to 56%, more than six times higher. Figure also used strict standards when determining
whether a task was successful. There was no partial credit for almost finishing the job.
For living room cleanup, every toy had to make it into the basket. For towel folding, the towels had
to be folded and placed into the basket. For bed making, the pillows and comforter had to reach
specified positions and the comforter had to be pulled into place. Another interesting part of
Helix 2.5 is its ability to recover when something does not go according to plan.
Figure demonstrated the robot stepping backward, changing its stance, repositioning itself, and even
moving around a bed to continue the task. That kind of correction becomes extremely important
when a robot is working for several minutes rather than performing one short, carefully scripted
movement. Figure also found that Helix 2.5 required two times less task-specific data
than a representative previous Helix 02 behavior while expanding the behavior across 30 unseen
homes. This could become important if humanoid robots are eventually expected to operate at large
scale. Training a robot separately for every house, warehouse, hotel, or workplace would be
extremely difficult. The company also reported evidence of a human to humanoid scaling law.
Researchers trained four models using progressively larger amounts of index data,
covering an eight times increase in pre-training data. Performance improved predictably as more
human experience was added. That does not mean household robotics has been solved.
A 56% full task success rate leaves plenty of room for improvement, especially if humanoid robots are
eventually expected to perform household work reliably every day. But the direction is important.
Instead of teaching a humanoid exactly how to operate inside one specific room,
Figure is trying to build a robot that carries useful physical knowledge from one environment
into another. For household humanoids, that could be one of the biggest steps toward making robots
genuinely useful outside laboratories and carefully prepared demonstration spaces.