We present Helix 2.5, a major update to our humanoid Vision-Language-Action model that achieves unprecedented zero-shot generalization across 30 unseen home environments. This represents a significant step toward deploying general purpose humanoid robots in real-world residential settings.
Building on the foundation of Helix 02, Helix 2.5 introduces improvements in whole-body control, self-correction capabilities, and data efficiency that together enable robust performance in environments the robot has never encountered during training.
Key Results
- 01Helix 2.5 generalizes to 30 unseen homes with zero additional training
- 02Index pretraining drives a 2x improvement in zero-shot generalization
- 03Helix 2.5 needs 2x less data than Helix 02 to reach the same performance
- 04Whole-body self-correction enables robust recovery from failures
- 05A human-to-humanoid scaling law predicts performance from human data
Video: Helix 2.5 in Action
The Whole-Body Generalization Problem
Deploying a humanoid robot in unseen environments requires generalization across multiple dimensions: novel room layouts, different object placements, varying lighting conditions, and diverse task configurations. Helix 2.5 addresses this through a combination of large-scale pretraining on the Index dataset and improved whole-body coordination.
Index Pretraining Drives Zero-Shot Generalization
The Index dataset, collected from thousands of human demonstrations across diverse home environments, provides the foundation for Helix 2.5's zero-shot capabilities. By pretraining on this large-scale dataset, the model learns robust representations that transfer to unseen homes without any additional fine-tuning.
Figure 1: Zero-shot generalization performance across 30 homes
Figure 1: Helix 2.5 achieves significantly higher success rates across all 30 evaluation homes compared to Helix 02.
Evaluating Three Whole-Body Behaviors in 30 Unseen Homes
We evaluate Helix 2.5 on three representative whole-body behaviors: navigation to target locations, object manipulation (pick and place), and multi-step task completion. Each behavior is tested across 30 previously unseen home environments.
Table 1: Success rates across three whole-body behaviors in 30 unseen homes.
Helix 2.5 Needs 2x Less Data
Despite achieving significantly better performance, Helix 2.5 requires approximately half the training data compared to Helix 02. This data efficiency gain comes from improved model architecture, better pretraining strategies, and the rich diversity of the Index dataset.
Whole-Body Self-Correction
A key qualitative improvement in Helix 2.5 is its ability to self-correct during task execution. When the robot encounters an unexpected obstacle or makes a suboptimal movement, it can recover and continue the task without human intervention. This capability is critical for reliable deployment in unstructured home environments.
A Human-to-Humanoid Robot Transfer Scaling Law
We discover a predictable scaling relationship between the amount of human demonstration data and the resulting humanoid robot performance. This scaling law allows us to predict robot capabilities from human data volume, providing a clear roadmap for future improvements.
Figure 2: Human-to-humanoid scaling law
Figure 2: Robot performance scales predictably with the amount of human demonstration data.
It's Time to Scale Up
Helix 2.5 demonstrates that with the right combination of large-scale pretraining, efficient model architecture, and whole-body coordination, general purpose humanoid robots can generalize to unseen home environments. The scaling law we've identified provides a clear path forward: as we collect more human demonstration data through the Index platform, robot capabilities will continue to improve predictably. The future of home robotics is here.