Figure AI Robot HELIX 2.5 Autonomous Generalization In 30 New Homes

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Figure AI Robot HELIX 2.5 Autonomous Generalization In 30 New Homes · Channel: AI News · 07:59 · Watch on YouTube

Topics: Task and workflow automation, AI models, Robotics and autonomous systems

Summary

Figure presents Helix 2.5 performing bed-making, towel-folding, and living-room tidying in 30 homes the robot had never encountered, with no house-specific data, fine-tuning, or adaptation. The video reports that an Index-pretrained policy succeeded in 56% of zero-shot trials, compared with 9% for a policy trained from random weights, and describes improvements in data efficiency and self-correction. Figure also highlights predictable gains from scaling pre-training and its investment in Index data generation and compute. The video then compares Figure’s generalization-focused approach with AGI Bot’s A3 Ultra, describing its deployments, hardware, capabilities, and safety features.

Key points

Helix 2.5 demonstrates three whole-body skills: tidying a living room, folding towels, and making a bed.
Figure says the skills were tested in 30 real homes without collecting data, fine-tuning, or adapting the robot in those homes.
Index pre-training reportedly raised zero-shot trial success from 9% for a random-weights policy to 56%.
Compared with Helix 02, Helix 2.5 reportedly matched its success rate with half as much task-specific training data and generalized across 30 homes instead of one.
Figure says Helix 2.5 can recover from mistakes by repositioning, changing stance, or walking around a bed.
Figure reports predictable performance improvements as pre-training data increased, and says its largest run’s test loss was forecast with 0.54% error.
Figure says Index generates roughly 35 minutes of new human-experience data per second and that the company has committed $3.5 billion in compute to Helix training.
The video describes AGI Bot’s A3 Ultra as deployed in settings including hotels, dealerships, supermarkets, and metro inspection routes.
The A3 Ultra is described as capable of autonomous charging or battery swapping, handling new tasks from a human demonstration, and coordinating with up to 100 units through the Link Bot system.

Chapters

Start	Chapter	Summary
00:00	Helix 2.5 and the challenge of new homes	The video introduces Helix 2.5 and asks whether it can perform useful household tasks in homes it has never seen.
01:04	Three household skills across 30 homes	Figure describes the bed-making, towel-folding, and living-room tidying tests and says they were conducted without prior data or adaptation in the homes.
03:08	Pre-training, efficiency, and self-correction	The video compares pre-trained and random-start policies, then covers reported gains in data efficiency, recovery from mistakes, and scaling behavior.
05:15	Figure’s broader claims and investment	Figure frames the results as evidence for transferring whole-body skills across environments and discusses Index data generation and compute commitments.
05:46	AGI Bot A3 Ultra	The video outlines the A3 Ultra’s deployments, specifications, task capabilities, safety features, and multi-robot coordination.

Quotes

My dream is to never have to make my bed again. — 00:00

In order to do that, we have to be able to deploy to any home. — 01:04

The random start policy succeeded on just 9% of zero-shot trials, but the Index pre-trained policy succeeded on 56% over six times higher. — 03:08

Over a long, unfamiliar task, Figure says Helix 2.5 will step back to reposition itself, change its stance, or walk around an entire bed to fix a bad fold — 04:13

Concepts and entities

Helix 2.5, Figure 3, Index, zero-shot generalization, whole-body generalization, Helix 02, Figure, AGI Bot, A3 Ultra, link bot

Full transcript (original language)

00:00 My dream is to never have to make my bed again. This is a home the robot’s never been in before. This one is important. Doing laundry is the worst. >> Is this embodied artificial general intelligence or is this? Today on the AI news, we’re taking a closer look at Helix 2.5 executing three new skills. Plus, we’ll see how well it generalizes like this. >> This is a home the robot’s never been in before. The robot’s never seen any of these objects. We’ve made no changes to any of

00:32 the training. All we’ve done is basically book this book this place, ship the robot here, and now it’s doing real useful work. And the robot’s constantly changing to the dynamics of the environment. From here, Figure 3 will go off and basically clean the entire living room. >> So, Figure just introduced Helix 2.5 as the most advanced neural network they’ve ever built for several reasons. But before you see all three autonomous skills from start to finish, Figure’s

01:04 pitch here is simple but hard to pull off. Can a humanoid generalize inside a home it’s never seen before to immediately get to work with zero prior data from that house? >> My dream is to never have to make my bed again. In order to do that, we have to be able to deploy to any home. What’s interesting here is Figure 3 has never been in this room before. It’s never seen this bed. It’s never seen this pillow. And it has to be able to do autonomous work fully end to end to make this bed.

01:35 This one is important. Doing laundry is the worst. At my house, we’re doing laundry every single day. This task is really difficult for robotics as it requires really precise manipulation. With Helix 2.5, we’re folding towels in a house the robot’s never been in with towels it’s never seen. >> And to test zero-shot generalization, Figure pre-trained Helix 2.5 on Index,

02:05 which Figure says is the largest robot training data set on Earth. Then produced three distinct whole-body skills from that single foundation model. Starting with skill number three, tidying a living room, the robot has to locate scattered toys, 13 to 15 of them per home, and place every single one in a basket. But this is significantly easier than skill number two, house folding. And while it sounds simple, the robot has never seen these specific towels, this specific folding surface, or this specific room before stepping

02:37 inside. Then there’s skill number one, which is the hardest of the three, making a bed. Both pillows and every comforter corner have to land at the top of the bed with the comforter pulled smooth on a bed frame the robot has never encountered. And Figure tested all three skills across 30 real homes with zero data collected in any of them beforehand, no fine-tuning, and no adaptation once the robot arrived. In fact, Figure says this is the first demo of zero-shot whole-body generalization at this scale on a humanoid robot. But

03:08 the more interesting question here isn’t whether it worked, but why. So Figure ran a direct comparison to find out. They trained two identical policies on identical task data, none of it from the evaluation homes. One started from random weights, the other started from the Index pre-trained Helix 2.5 foundation. Everything else, including architecture, training, evaluation, it all stayed fixed. The random start policy succeeded on just 9% of zero-shot trials, but the Index pre-trained policy succeeded on 56% over six times higher.

03:42 And since pre-training was the only variable that changed, that gap is a direct measurement of what broad pre-training on human behavior actually buys. And here’s where it gets even more efficient. Comparing Helix 2.5 to a Helix 02 policy trained on the same task, Helix 2.5 matched its predecessor’s success rate using half as much task-specific training data, then generalized that behavior across 30 homes it had never seen instead of just the one environment it was trained in. So, Figure frames this as behavior

04:13 specification getting twice as cheap while its usable scope expanded 30-fold. But, there’s also a qualitative shift worth flagging here, self-correction. Over a long, unfamiliar task, Figure says Helix 2.5 will step back to reposition itself, change its stance, or walk around an entire bed to fix a bad fold, recovering from mistakes rather than getting stuck on them. And Figure credits this ability directly to Index pre-training, not task-specific practice. And then there’s the scaling law, because Figure trained four

04:45 versions of the model on progressively larger slices of Index. An eight-fold increase in pre-training data while holding model size and downstream training constant, and the model’s performance improved predictably with each doubling, precisely enough that Figure could forecast their largest run’s test loss to four decimal places before ever running it, with a forecasting error of just 0.54% across the entire range. And that’s the same kind of predictable scaling behavior that transformed large language

05:15 models, only now Figure claims it’s on a humanoid robot for the first time. More importantly, Figure argues this is the first real evidence that whole-body intelligence can be learned broadly from human experience and then transferred to new places instead of just being rebuilt separately every single time a robot enters a new environment. And the scale behind this is substantial. Figure says Index is now generating roughly 35 minutes of new human experience data every second, and the company has committed $3.5 billion in compute

05:46 specifically to training Helix going forward. So, tell us in the comments below how close is this to actual embodied general intelligence? And how does it stack up to this brand new robot, the A3 from AGI Bot? >> Because if figures betting on generalization, AGI Bot is working towards a similar direction, but with a twist. So, AGI Bot’s A3 Ultra just moved from mass production into actual deployment, already working shifts in hotels, auto dealerships, supermarkets, and metro inspection routes.

06:19 And as for specs, it stands at 174 cm tall and weighs 60 kg, with the A3 being able to autonomously dock itself to charge or even swap out its 1,152 Wh battery in just 10 seconds flat, rather than waiting through a full 2-hour charge, which is how it stretches that 8-hour battery rating into 24-hour coverage. And as for intelligence, its 700 tera operations per second onboard compute handles perception and path planning in real time, backed by lidar

06:53 and full 360° surround view cameras, allowing it to navigate a lobby or a supermarket floor while tracking people and obstacles simultaneously. Plus, its 20° of freedom tactile hands can handle anything from a single sheet of paper to up to a 5 kg object, and the system can even pick up an entirely new task just from one human demo rather than needing to be explicitly reprogrammed. And it’s already working around actual customers and staff, with the A3 Ultra leaning on a four-layer safety system, predicting risk 3 m out, stopping on contact,

07:27 running dual emergency stops, and catching itself if it starts to fall. And it’s all validated through collision, drop, and extreme temperature testing. And when the job calls for more than one robot, AGI Bot’s link bot system can coordinate up to 100 units at once. Plus, the unbox and deploy of the A3 is pretty cool. So, since there isn’t any pricing available publicly, tell us in the comments whether you think this would be more or less expensive than the Figure 3 and which one you’d choose. Anyways, like and subscribe and check out this video here for more of the latest in AI and robotics news. And

07:59 thanks for watching.