HyperAIHyperAI

Command Palette

Search for a command to run...

AGI

Figure AI Finalizes Figure 04 Design, Reveals Key Next-Gen Model Progress

Figure AI used an approximately one-hour all-hands presentation to outline plans for next-generation humanoid robots, data collection and compute capacity. Founder and CEO Brett Adcock said Figure 04 has completed design freeze and entered component manufacturing and lab testing, including actuator and battery verification. He described the robot as potentially the first iPhone moment for humanoid robotics, with a dexterous hand redesigned four times to track fingertip and contact positions near millimeter precision. The fourth generation follows Figure 03, which has been produced beyond 1,000 units after a more than 90 percent cost reduction. Figure AI also detailed the Helix software stack. Helix 2.5, released in September, showed broader generalization in 30 unseen San Francisco Bay Area homes, improving success rate from 9 percent for baseline training to 56 percent without environment-specific fine-tuning. The company attributed the result to Index pretraining data, a global crowdsourced video program launched in August. Adcock said Index has collected 37.5 million videos, 1.26 million app downloads and contributions from 112 countries, growing about 20 percent weekly over 10 weeks and adding roughly 116,600 hours of video daily. Data spending reached 8.1 million dollars in the latest week, which he said could exceed the total funding raised by many robotics firms over the next 12 months. To process this data, Figure AI said its Nscale collaboration involves an initial 3.5 billion dollar commitment and potential expansion to about 6 billion dollars for up to 100,000 Nvidia Vera Rubin GPUs. Infrastructure is under development in Texas, with power permits and procurement underway and operation targeted for 2027. Figure plans to double current compute in the near term. Adcock argued robotics pretraining may require larger clusters than language models because robots must process continuous video, spatial information and complex motion. The company is scaling manufacturing through BotQ, which reached production of one Figure 03 per hour from one per day in under 120 days. It has deployed robots in commercial facilities in three US states, including a logistics test that ran about 200 hours over eight days by rotating charged units, and an industrial trial in an air-conditioning-free Los Angeles facility at about 40 Celsius. Adcock said Figure holds more than 2 billion dollars in cash to support data, compute and manufacturing expansion. The presentation concluded with a teaser for an unannounced model trained on the company's largest pretraining dataset and GPU allocation. Adcock described an internal Black Box test weeks earlier as among the most important moments in his robotics research career and called the early result Baby AGI, while saying the team is training a larger version with updates expected in the coming weeks. Figure's strategy is to apply scaling laws, similar to large language models, by expanding human action data, compute and hardware precision to make robots more general-purpose.

Related Links