Figure unveils Helix, its own vision-language-action model that drives a humanoid's entire upper body from two onboard GPUs

Figure AI, the US humanoid developer, on 20 February 2025 published Helix, a vision-language-action (VLA) model that it says controls the full upper body of its humanoid robot from camera images and plain-language commands. In a technical post on its website, the company described Helix as a single network that lets its robots pick up household objects they have never seen and lets two robots share one task.

The numbers Figure disclosed are specific. Helix splits into two parts. System 2, the slower layer, is a 7-billion-parameter open-source, open-weight vision-language model pretrained on internet-scale data and running at 7 to 9 Hz. System 1, the fast layer, is an 80-million-parameter cross-attention encoder-decoder transformer that runs at 200 Hz. Together they output commands for a 35-degree-of-freedom action space covering wrist poses, finger flexion and abduction, and torso and head orientation targets.

Figure said Helix was trained on roughly 500 hours of teleoperated robot behaviour, which it put at less than 5% of the size of previously collected VLA datasets. To pair those recordings with language, the company ran an auto-labelling vision-language model over video clips and asked it, in effect, what instruction would have produced the action on screen. Training ran end to end, from raw pixels and text to continuous actions, with a standard regression loss, a single training stage and a single set of weights, according to the post.

The demonstrations centred on a kitchen. In Figure's videos two robots unpack a bag of groceries they have not handled before and put the items away, passing objects to each other on spoken prompts such as handing a bag of cookies to the robot on the right. Figure called Helix the first VLA to run simultaneously on two robots and the first to output high-rate continuous control of an entire humanoid upper body. It also said its robots could now pick up thousands of novel items in clutter, from glassware and toys to tools and clothing, and that all objects used in training were excluded from the evaluations. These are the company's own claims and have not been independently tested.

The hardware choice may matter as much as the model. Figure said Helix runs entirely onboard on dual low-power embedded GPUs, with System 2 and System 1 each on its own processor, and described it as the first VLA to do so and therefore ready for commercial deployment. Six days later, on 26 February 2025, Figure published a logistics follow-up stating that a Helix policy trained on 8 hours of demonstrations reached an effective throughput above 1.1, or 10% faster than the human-led demonstrations it learned from, after adding stereo vision, multiscale visual features, learned self-calibration and a test-time speed-up it calls Sport Mode.

Figure enters this field with a large balance sheet and a notable history. On 29 February 2024 it announced a $675 million Series B at a $2.6 billion valuation, with investors including Microsoft, the OpenAI Startup Fund, NVIDIA and Bezos Expeditions, and said it had signed a collaboration agreement with OpenAI to develop next-generation AI models for humanoid robots. The Helix post does not mention OpenAI and presents the model as Figure's own work. In September 2025 Figure went on to raise a $1 billion Series C at a $39 billion post-money valuation, according to the company.

Helix lands in a crowded race to build robot foundation models. Physical Intelligence, a San Francisco start-up, unveiled π0 on 31 October 2024, a model built on a 3-billion-parameter vision-language model that uses flow matching to issue motor commands at up to 50 times a second, trained across eight robot configurations. The company has since published π0 code and weights under an Apache-2.0 licence in its openpi repository, which states that its base checkpoints were pre-trained on more than 10,000 hours of robot data. Google DeepMind followed on 12 March 2025 with Gemini Robotics and Gemini Robotics-ER, built on Gemini 2.0, which it said more than doubled performance on a comprehensive generalisation benchmark against other state-of-the-art VLAs, with Apptronik as its humanoid partner.

The architecture follows an idea long used in robotics and borrowed from psychology. A slow, deliberative layer interprets the scene and the instruction, then hands a compact summary to a fast reflexive layer that steers fingers and wrists. In Helix that summary is a single continuous latent vector, which System 1 combines with its own visual features from a fully convolutional vision backbone that Figure says was pretrained entirely in simulation. The split lets the large language-aware model think at a speed it can sustain on embedded hardware, while the small policy keeps the arms and hands smooth at 200 Hz.

For the global market the significance is that a hardware company with a production plan now claims to own its whole software stack, rather than renting intelligence from a model lab. Open questions remain. Figure has not named the open-weight model inside System 2, has not published success rates for the grocery demonstration, and has not given the GPU model or power draw. Helix as described controls the upper body only, leaving locomotion to other systems, and 500 hours of data is small next to the more than 10,000 hours Physical Intelligence cites.