Releases·Model Releasefeatured
Figure’s Helix 2.5 Generalizes Household Robot Tasks Zero-Shot Across 30 Homes
Figure AI announced Helix 2.5, a vision-language neural model trained on millions of examples of household work and deployed in its humanoid robots. The company reports zero-shot generalization of household tasks across 30 unseen home environments and published 4-hour footage of robots performing tasks in homes. Figure says it has committed $3.5B of compute to training Helix and generates new data at 35 hours per second, signaling a push toward physical intelligence in home robotics.
390 engagement3 sourcesThu, Sep 17, 2026, 07:05 PM
AI Overview
- Figure ships Helix 2.5, which executes household robot tasks zero-shot in 30 homes without per-home training
- Figure announced Helix 2.5 on 2026-09-17 via a Reddit post drawing 71 engagements
- Figure generates data at 35 hours per second for Helix training and has committed $3.5B of compute to training Helix
- Figure trained Helix 2.5 on millions of examples of household work and released 4-hour footage of the robots operating across homes
Related
Social·Community
Figure's Helix 2.5 claims zero-shot generalization across 30 homes; critics cite ~44% task failure
Figure released Helix 2.5, showcasing zero-shot performance across 30 unseen Bay Area homes on tasks like folding towels and tidying rooms. A widely shared critique argues the demo overstates generalization: the robot reportedly fails ~44% of full-task trials on tasks it was specifically trained on, and the showcased homes are unusually clean and uncluttered.
Figure released Helix 2.5, showcasing zero-shot performance across 30 unseen Bay Area homes on tasks like folding towels and tidying rooms. A widely shared critique argues the demo overstates generalization: the robot reportedly fails ~44% of full-task trials on tasks it was specifically trained on, and the showcased homes are unusually clean and uncluttered.
45 engagement1 sourceredditSat, Sep 19, 2026, 08:05 AM
Products·launched
Figure humanoid robot working untrained in 30 new homes, shown in 4-hour video
Figure.ai released a long-form video showing its humanoid robot performing tasks in 30 homes it had not been trained for. The robot presumably relies on large multimodal models for perception and instruction following, though the posts do not specify the model stack. The demo targets home robotics, showing generalization without per-home training.
Figure.ai released a long-form video showing its humanoid robot performing tasks in 30 homes it had not been trained for. The robot presumably relies on large multimodal models for perception and instruction following, though the posts do not specify the model stack. The demo targets home robotics, showing generalization without per-home training.
11 engagement1 sourceredditFri, Sep 18, 2026, 05:03 PM
Products·building
Figure.AI demos F.03 climbing a ladder autonomously
Figure.AI is demonstrating its humanoid robot F.03 autonomously climbing a ladder, showcasing advanced locomotion and manipulation capabilities. The demo highlights progress in real-world robot autonomy, likely leveraging AI models for perception and control. This addresses the challenge of deploying robots in environments requiring vertical mobility.
Figure.AI is demonstrating its humanoid robot F.03 autonomously climbing a ladder, showcasing advanced locomotion and manipulation capabilities. The demo highlights progress in real-world robot autonomy, likely leveraging AI models for perception and control. This addresses the challenge of deploying robots in environments requiring vertical mobility.
603 engagement1 sourceredditSat, Aug 1, 2026, 09:04 PM
Releases·Model Releasefeatured
Google DeepMind launches Gemini Robotics 2 for whole-body robot control
Google DeepMind released Gemini Robotics 2, a multimodal AI system that combines a vision language model and two vision language action models to control various robots, including humanoids. The system enables whole-body motions and dexterous tasks like screwing in lightbulbs and tying trash bags. One sub-model is publicly available for developers.
Google DeepMind released Gemini Robotics 2, a multimodal AI system that combines a vision language model and two vision language action models to control various robots, including humanoids. The system enables whole-body motions and dexterous tasks like screwing in lightbulbs and tying trash bags. One sub-model is publicly available for developers.
250532 engagement8 sourcesrss, redditThu, Jul 30, 2026, 03:02 PM
Products·launched
Sunday Robotics ACT-2 model achieves 99% success on zero-shot laundry folding
Sunday Robotics released ACT-2, a model that enables a robot to fold laundry with 99% success on unseen items. It uses LLMs for task understanding and planning, solving the problem of robotic dexterity in unstructured home environments for consumers.
Sunday Robotics released ACT-2, a model that enables a robot to fold laundry with 99% success on unseen items. It uses LLMs for task understanding and planning, solving the problem of robotic dexterity in unstructured home environments for consumers.
126 engagement1 sourceredditFri, Jul 17, 2026, 05:28 PM