Skip to main content

Figure's Helix 2.5 Boosts Robot Success Rate to 56%

Figure's Helix 2.5: A Robot That Learns Like a Human

Imagine a robot walking into a stranger's home and successfully tidying up—without any prior training on that specific environment. That's exactly what Figure's latest neural network, Helix 2.5, achieved in a recent test.

On September 17, Figure unveiled Helix 2.5, a humanoid robot brain pre-trained on the company's Index human behavior dataset. The real-world trial took place in 30 households across the San Francisco Bay Area—homes the robot had never seen before. The result? A zero-shot success rate of 56%, a massive leap from the 9% achieved by strategies trained from scratch under identical conditions.

What Can It Do?

Helix 2.5 is built on a single Index pre-trained base model and handles three full-body tasks: organizing the living room, folding towels, and making the bed. Crucially, the evaluation didn't include any task-specific data for these homes, and success required complete task completion. The robot even showed full-body self-correction—like stepping back to adjust its position, changing stance, or moving across a bed to finish the job.

The Power of Pre-training

To isolate pre-training's impact, Figure kept everything else constant: model architecture, optimization, hyperparameters, downstream data, and evaluation conditions. Only the initialization method changed. The company reports that no single evaluation task in the Index dataset accounted for more than 1.90% of the data, meaning the results directly reflect pre-training's contribution to generalization. Doubling the Index data led to a predictable drop in action prediction loss—a sign that data scaling laws apply to humanoid robots too.

Index: The Data Engine Behind It

Index launched on August 25 and has already racked up 264,000 downloads, over 44,000 weekly active users, and more than 16 million videos. Figure has paid $15 million to data contributors and invested $3.5 billion in computing resources to train the Helix series. With Helix 2.5, the company also achieved comparable success rates using only half the adaptive data of the previous Helix02.

Key Points

  • Zero-shot success rate jumped from 9% to 56% in 30 unseen homes.
  • Three full-body tasks: organizing living room, folding towels, making bed.
  • Pre-training on Index dataset is the key differentiator, with no task-specific data.
  • Data scaling law observed: doubling Index data reduces action prediction loss.
  • Index dataset has 264k downloads, 44k weekly active users, and 16M+ videos.
  • $3.5 billion invested in computing for Helix training.

Image

Image