Humanoid robot company Figure has announced the release of its latest neural network model, Helix 2.5. The team rented 30 previously untested residential units in the Bay Area, and the robots began performing practical household tasks without any additional targeted training upon arrival.

Helix 2.5 is pre-trained on Figure's "global-scale human behavior dataset" called Index. A foundational model pre-trained on Index has been adapted to three distinctly different behavioral tasks, covering capabilities such as moving, rigid and flexible object manipulation, dual-arm coordination, and active perception. The company stated, "Transferring task-specific data from a model already trained on extensive human experience yields significantly better results."

The specific performance data for the three full-body tasks tested are as follows: organizing the living room (picking up 13 to 15 scattered toys and placing them in a basket) had a success rate of about 40% (56 out of 140 attempts); folding towels and placing them in a basket had a success rate of about 62% (87 out of 140 attempts); making the bed (arranging pillows and blankets and smoothing the bed surface) had a success rate of about 67% (94 out of 140 attempts). In total, there were 420 attempts across the three tasks, with 237 successes, resulting in an overall zero-shot success rate of 56%.

In comparison, a baseline model trained from scratch without pre-training on Index data achieved a success rate of only 9% under the same testing conditions. Meanwhile, Helix 2.5 achieved performance comparable to Helix 02 using only half the amount of task-specific data from Helix 02, and this was accomplished in 30 unfamiliar residences where no prior data had been collected. Helix 02 previously required data collection specific to the environment to achieve usable results.

The Index dataset supporting this capability leap was officially released on August 25. This task was completed using the current generation humanoid robot platform Figure 03, equipped with the Helix 2.5 software layer. Figure CEO Brett Adcock referred to Helix 2.5 as "the most important project in Figure's history."

In terms of financing background, Figure completed a $1 billion Series C funding round on September 16, 2025, led by Parkway Venture Capital, achieving a post-money valuation of $39 billion. Investors included NVIDIA, Baillie Gifford, Intel Capital, Macquarie Capital, Salesforce, T-Mobile Ventures, and Qualcomm Ventures. The company plans to ship 100,000 humanoid robots over the next four years while expanding its Helix embodied intelligence platform and the BotQ large-scale manufacturing facility's capacity.

From a market mechanism perspective, the increased zero-shot testing success rate across 30 households is seen as key evidence to demonstrate to potential customers and investors that its technology can be scaled without needing to collect data for each new scenario, directly supporting the narrative of scalability behind the company's $39 billion valuation. For existing investors like NVIDIA and Qualcomm, the accelerated commercialization of humanoid robots suggests a further realization of demand for computing power and component orders in the embodied intelligence sector. For competitors in the same space, such as Tesla's Optimus, Figure's replicable data in real home environments will intensify competition regarding "who can achieve large-scale deployment in home scenarios first."

Source: Public Information