Figure 03 vs Figure 02: Upgrades, Evidence and Limits

2 October 2026

Figure AI’s Figure 03 humanoid robot changes sensing, charging and manufacturing. Their value varies by job: factory production, logistics demonstrations and household tests offer different levels of proof.

Figure 03 launched on 9 October 2025. The recent news is Figure’s 30 September 2026 retirement announcement: most Figure 02 robots are decommissioned; a few remain at headquarters. The announcement mentions future Figure 04 development without launching it or releasing specifications.

What changed in Figure 03?

Area	Figure 02	Figure 03
Hardware: vision	Earlier camera system	Figure reports twice the frame rate, one-quarter the latency and a 60% wider field of view per camera.
Hands and sensing	Earlier hand design	Palm cameras and fingertip tactile sensors support manipulation.
Charging and body	Mass comparison baseline	9% less mass; softer, washable coverings; revised audio; 2 kW inductive charging through the feet.
Manufacturing	Relied heavily on CNC machining	Molded, cast and stamped parts; Figure reports lower build cost.
Demonstrated work	Sheet-metal loading in BMW production	Parts sequencing and cart pulling in a BMW logistics demonstration.

Hardware comparisons are Figure’s reported changes. Work examples come from BMW’s project account and Figure’s demonstration. These sources provide no controlled, same-task comparison between the robots.

What BMW’s deployments show

BMW’s robotics projects offer the clearest evidence of sustained work by Figure’s humanoids. BMW’s 25 June 2026 update says Figure 02 helped produce more than 30,000 BMW X3 vehicles over ten months. The robot inserted sheet-metal parts for welding, demonstrating a defined production task, not entire-car assembly.

The Figure 03 project targets sorting incoming components into a sequencing trolley, then moving it to a collection point. BMW describes both projects here.

A higher camera frame rate cannot predict usable parts handled per shift. That requires measuring the intended workflow, including failed grasps, interruptions and recovery. BMW’s account supports the earlier deployment and next project’s scope, without establishing that Figure 03 does Figure 02’s job faster.

Figure AI’s 30 June 2026 demonstration shows Figure 03 coordinating its hands, arms, torso and feet with Helix 02 while picking, sequencing and pulling carts.

Figure describes the demonstration here.

What Helix 2.5 adds

Figure 02 and Figure 03 are hardware generations; Helix 02 and Helix 2.5 are software versions. Do not frame a software evaluation as a hardware comparison.

On 17 September 2026, Figure reported testing Helix 2.5 on tidying toys, folding towels and making beds in 30 unseen Bay Area homes. Its Index-pretrained model achieved 56% whole-task success, versus 9% for a randomly initialized model. The comparison kept task setup and other training conditions fixed.

It tested pretraining’s contribution, not Figure 02 versus Figure 03. These are manufacturer-reported results, not independent validation. Source: Figure’s Helix 2.5 evaluation.

“Zero-shot” refers to unseen homes and objects; task training occurred elsewhere. The 56% completion rate shows some transfer to unfamiliar settings but substantial reliability gaps. It establishes neither dependable, complete housework nor success on unevaluated chores.

Battery life, cost and availability

Figure’s 17 July 2025 battery report specifies 2.3 kWh and claims up to five hours at stated peak performance. Runtime varies by workload; treat the claim as a planning input, not a guaranteed shift. This evidence supports no Figure 02 runtime comparison. Trials should measure productive time, charging downtime and actual route and payload demands.

Lower manufacturing cost does not establish retail pricing; the launch announcement provides no public retail price. These sources establish neither general consumer availability nor delivery commitments. Buyers should request commercial terms before budgeting.

Assess autonomy, supervision and workload: who handles failed tasks, how often help is needed, and what constitutes completion. For home robot planning, read Rohumn’s home humanoid buying guide and home robot safety checklist. Evaluate Figure 03 against a defined job; complete housework readiness remains unproven.