Figure AI - Helix 2.5 Robot: Zero-Shot Home Generalization (www.figure.ai)

Figure AI has announced the Helix 2.5 robot, marking a significant advancement in robotics with its ability to perform multiple household tasks in completely unfamiliar environments without prior data collection or fine-tuning. This humanoid robot showcases "zero-shot whole-body generalization," successfully executing long-horizon tasks like tidying living rooms, folding towels, and making beds across 30 distinct homes. The robot was pretrained on the Index dataset, which captures a wide array of human behaviors, enhancing its capability to adapt to new surroundings and objects while requiring only half as much task-specific data compared to previous models, such as Helix 02. This breakthrough is pivotal for the AI and ML community as it indicates that robots can acquire generalizable behavioral skills from human experiences rather than needing to relearn in every new environment. The substantial increase in task success rates—from 9% to 56%—demonstrates the effectiveness of Index pretraining, reinforcing the hypothesis that broad pretraining can streamline robot functionality across diverse scenarios. Helix 2.5's performance not only sets a new bar for humanoid robotic capabilities but also hints at the potential for scaling data use in robotics, suggesting a transformative path forward in achieving general physical intelligence.