@rohanpaul_ai: Figure may have found a scaling rule for robot pretraining: keep the model and task training fixed, add more Index data…
Summary
Figure has discovered a scaling rule for robot pretraining where adding more human behavior data predictably reduces action-prediction loss, similar to LLM scaling principles. Their Helix 2.5 model achieved over sixfold improvement in zero-shot household task success across unseen environments.
View Cached Full Text
Cached at: 09/18/26, 10:38 AM
Figure may have found a scaling rule for robot pretraining: keep the model and task training fixed, add more Index data (Figure’s large pretraining dataset of human behavior)
and action-prediction loss falls in a clean, predictable way.
the smaller runs predicted the 8x-data run almost exactly, i.e. a robotics company can estimate what another doubling of human-behavior data will buy before spending the compute on the full run.
should make robot training less trial-and-error and more like LLM scaling, although the curve predicts action loss, not real-world task success.
so it does not yet prove equally predictable gains in robot reliability.
Rohan Paul (@rohanpaul_ai): Figure just released this video. a beautiful robot future is indeed coming.
Its Helix 2.5 model (Figure’s end-to-end robotics brain) lifted zero-shot household-task success more than sixfold across 30 unseen homes.
Figure also reports success rising from 9% to 56%, with
Similar Articles
Figure’s Helix 2.5 Generalizes Household Robot Tasks Zero-Shot Across 30 Homes
Figure releases Helix 2.5, an AI model that enables robots to generalize household tasks zero-shot across 30 homes without additional training.
Does Scaling Web-Video Pre-training Help Real Robots Do Real Work? (36 minute read)
This study rigorously tests whether scaling web-video pre-training improves robot task performance, finding that larger models and more compute consistently yield better policies on real industrial tasks, with pre-training quality correlating with performance.
@HuggingPapers: Scaling video pre-training to 120K hours ZimaBlue frames video scaling as a route to generalizable World Action Models.…
Scaling video pre-training to 120K hours boosts zero-shot success in World Action Models from 36.1% to 77.8% on real robots, enabling faster action prediction.
HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining
This paper finds that egocentric human video, when processed with a filtering and labeling pipeline, can outperform teleoperated real-robot data for pretraining embodied foundation models, achieving lower validation loss and higher success rates on real-robot tasks.
Figure.ai just released a full 4-hour video showcasing their humanoid working in 30 new homes without requiring any additional training
Figure.ai released a 4-hour video demonstrating their Helix 2.5 humanoid robot working in 30 new homes without any additional training, showcasing its generalization capabilities.