Figure 02 — the humanoid building for factory work
Figure's pitch isn't a research demo. It's "this robot will work in a BMW plant in 2026." Here's what the hardware looks like and what they got right.
Figure 02 is the second generation of Figure AI's humanoid. The Gen 1 was a research platform; Gen 2 is the version they're shipping to industrial pilots. BMW is the public design partner — Figure 02 robots are now in BMW's Spartanburg plant doing actual work (mostly part-handling on the body-in-white line).
What changed from Gen 1
Battery — moved into the torso, doubled capacity, runtime now ~5 hours of factory work.
Hands — 16 degrees of freedom per hand (up from 6 on Gen 1). Each finger is independently actuated. This is where most of the engineering went, because manipulation is the gating factor for "useful humanoid in a factory."
Compute — onboard NVIDIA Jetson Thor. The visual policy network runs at 200 Hz on-device. No cloud round trip — that wouldn't pass a factory's network/safety review anyway.
Communications — speakers and microphones in the head. The robot can talk to workers and respond to spoken commands. OpenAI's GPT-4-class model handles the language layer.
The strategic bet
Figure's bet is that humanoids will become economically useful before "general purpose". They're targeting factories with very specific workflows that already have ergonomic and safety problems for humans (lifting, repetitive arm motions). The robot doesn't need to do everything; it needs to do one task per shift, reliably, at the speed of a human worker.
This is different from Tesla Optimus, which is positioned more broadly toward "household + everywhere." Figure is much narrower and consequently easier to evaluate: either the BMW pilot extends or it doesn't.
What's still missing
No humanoid today, including Figure 02, can match human dexterity for arbitrary objects. The robots demonstrated picking up a Tupperware lid, a battery pack, a coffee cup — all from known locations. They cannot, yet, reliably handle a tangle of cables or a deformable bag.
The training pipeline is still teleoperation-heavy. A human operator does the task many times in VR; the robot's policy network learns from those demonstrations. True generalization to unseen tasks remains an open research question.
Funding
Figure raised $675M at a $2.6B valuation in early 2024, backed by Microsoft, NVIDIA, OpenAI, Jeff Bezos. By any historical robotics benchmark, that's a lot of capital betting on humanoid factory work landing.
Compare with Tesla Optimus and our other Lens piece on Tesla's walking demo.
Ask R2 Co-pilot anything you didn't understand. It'll explain it plainly.
Tesla Optimus walking — the 4-minute version of a 22-minute reveal
Tesla's Optimus reveal video was 22 minutes long. The actually-new robotics in it fits in 4 minutes. Here's wh…
4:12 · HumanoidHow a Roomba decides where to clean
It's not random. A modern Roomba runs SLAM, builds a map of your home, and plans a route. Here's what's inside…
3:48 · ConsumerThe DJI drone gimbal in slow-motion — three motors, 8,000 corrections per second
That impossibly steady drone shot you've seen on Instagram? It's a three-axis robotic arm fighting gravity 8,0…
5:30 · Drones