A Billion-Dollar Bet, Paid Off in Four Weeks
Humanoid robotics startup Figure AI is back in the spotlight. Weeks after committing billions to building robot training data, the company has delivered: its next-generation neural network, Helix 2.5, achieved a 56% success rate on household chores across 30 Bay Area homes its robots had never seen — a 6X jump over the 9% baseline of a model trained from scratch, and a milestone for zero-shot generalization in humanoid robotics.
30 Unseen Homes, 420 Attempts
On September 17, Figure unveiled Helix 2.5 alongside an unusually honest test. The company rented 30 Bay Area homes and sent in robots that had never entered them, working on objects — toys, bedding, towels — that never appeared in the training data, all using one single frozen checkpoint of identical weights.
Broken down by task, the picture is more nuanced: bed making succeeded 67% of the time (94 of 140 attempts), towel folding 62% (87/140), and tidying toys back into a basket lagged at just 40% (56/140). The grading was strict — full completion was required for credit, a crooked pillow meant failure, and every safety intervention counted against the robot rather than as a gray area. The final tally: 237 of 420, or 56%.
The Secret Weapon: the Index Data Platform
The driving force behind the leap is Index, the data product Figure took out of stealth on August 25 — what the company calls the most diverse robot training dataset ever assembled. Index collects roughly 35 minutes of human experience footage every second, uploaded by tens of thousands of weekly contributors through an app. With Index pretraining, Helix 2.5 needed 50% less task-specific data than its predecessor Helix 02, yet delivered stronger results across three times as many homes.
CEO Brett Adcock called it “the most important project we’ve ever taken on at Figure,” describing zero-shot generalization — a humanoid walking into a house it has never seen and getting straight to work — as the holy grail of the industry.
Is 56% Good Enough? Rivals and the Money Race
Whether 56% is impressive depends entirely on your yardstick. Against a from-scratch model, it is a six-fold jump. Against the bar for a robot you would trust alone in your living room, toy tidying remains a task it fails more often than it succeeds. Competitor Sunday Robotics recently reported a 99.1% success rate (778 of 785) for its ACT-2 system folding laundry in unseen environments — though the two results aren’t directly comparable given different tasks, grading rules, and definitions of “unseen.”
On the money side, Figure closed more than $1 billion in Series C funding and struck a Brookfield partnership earlier this month, giving it — by Adcock’s own account — the strongest balance sheet in humanoid robotics, with the home-robot timeline moved up by two years.
Same Week: Qualcomm Buys PickNik as the Physical AI Race Heats Up
In the same week as Figure’s announcement, Qualcomm said on September 22 it would acquire PickNik, the robotics software company that stewards MoveIt, one of the most widely used open-source robot manipulation frameworks. Qualcomm plans to integrate MoveIt with its Dragonwing robotics platforms and Arduino, and has pledged to keep MoveIt open source. Qualcomm shares dipped about 1.15% in premarket trading as markets read the deal as a long-term physical AI play.
Conclusion: Physical AI Enters the Data-Is-King Era
Figure’s 6X leap validates an increasingly clear industry logic: in physical AI, pretraining on massive general human behavior beats narrow task-specific practice — an echo of the scaling laws that transformed large language models. With Index’s 35-minutes-per-second data firehose, a billion dollars of runway, and chip giants like Qualcomm rushing to secure their positions, humanoid robots are moving from the era of choreographed demo videos to the era of raw numbers. A robot in every home is still some way off, but the direction is unmistakable: whoever owns the data owns the ticket to the next era.
