Figure AI Market Research

1. Introduction

Figure AI's point of differentiation within this group is the "high valuation + automaker scenario": it is the most highly valued independent humanoid robotics company in the group (Series C at $39B), and also the first vendor to complete a long-term pilot on a leading automaker's production line and enter commercial delivery in retail logistics. Unlike Unitree (hardware at scale) and Agibot (full-stack open source), Figure follows a deeply bundled route of "own humanoid + self-developed Helix model + RaaS commercialization."

1.1. Company Overview

1.2. Development History and Funding Milestones

1.3. Business Model and Pricing

Figure's business model centers on RaaS (Robot as a Service) as its public narrative: robots are charged to customers by running hours rather than sold outright. A third-party wiki source discloses that the BMW deployment is priced at about $25 / robot running hour (this figure is Credibility Level C, for reference only); official RaaS pricing is undisclosed, marked [To be filled]. The Project Go-Big partnership with Brookfield instead collects first-person human video in residential settings under a "data-for-deployment" model, belonging to a data-strategy type of collaboration.

2. Glossary

3. Feature Description and Product Line

3.1. Hardware Product Line

Figure's in-service models are the Figure 02 (the main BMW pilot unit) and the Figure 03 (released 2025-10-09). The Figure 03 is 9% lighter than the 02, uses a soft-coated material, adds a palm camera and tactile fingertips (3g force detection), and supports wireless charging; perception is upgraded to 2x frame rate, 1/4 latency, and a 60% wider field of view. Whole-machine specs: 1.68m / 60~61kg / 20kg payload / 1.2m/s / 5-hour battery life.

3.2. Helix and System 0

1. Helix long-horizon operation: Helix 02 has demonstrated 8 hours (later extended to 40 hours) of fully autonomous package sorting; in 2026-05 it completed a 200-hour marathon — about 250,000 packages, zero failures.
2. Long-horizon home tasks: dishwasher loading/unloading, bottle-cap twisting, and complex movements coordinating feet and hips, reportedly 100% completed without teleoperation.
3. Multilingual interaction: conversed in 11 languages at the White House AI summit.
4. System 0 (2026): perception-conditioned whole-body control that autonomously climbs stairs and traverses slopes and uneven terrain using only a binocular stereo camera, officially described as zero-shot Sim2Real.

4. Platform Architecture

4.1. End-to-End VLA Architecture

Figure's architectural core is Helix: perception, language understanding, and action generation are all handled by a single end-to-end model, rather than a traditional "perception—planning—control" modular pipeline. System 0 stacks a perception-conditioned whole-body control layer on top, enabling the humanoid to overcome obstacles on two legs without relying on heavy sensors such as LiDAR. Compared with Agibot (models + world model + simulation + OS, all open source) and PI (pure model licensing), Figure's architectural philosophy is deep coupling of model and humanoid, delivered closed-source.

4.2. BotQ Factory and the Data Flywheel

BotQ is Figure's lever for vertical integration: on one hand it ramps at a pace of 1 unit / hour (about 55 units / week) (nameplate 12,000 units / year, with a goal of 100,000 units in 4 years); on the other, Adcock disclosed in 2026-06 that the number of deployed robots has surpassed the number of employees — every deployed Figure 03 is a data-collection node, forming a "deployment → data → model → more deployment" flywheel. This corresponds to L4 (data assets) and L5 (real-world-condition evaluation) of the six Harness layers.

4.3. Commercialization Milestones Timeline

Figure 3-1｜Figure AI funding and commercialization milestones (2023—2026)

Data sources: prospectus-grade public coverage and third-party company profiles, information as of 2026-09-12; the BMW return date has three inconsistent versions, detailed in Section 6.1.

5. Harness Design

5.1. L1 Context Engineering Layer

Helix end-to-end assembles visual and language inputs into action outputs; under the VLA paradigm, "context engineering" is internalized in the model itself. The first-person human video collected by Go-Big is context construction at the pretraining-corpus level. Figure has not disclosed explicit context components such as retrieval/compression, so it is rated as implicitly carried within the model.

5.2. L2 Tools and Execution Layer

Figure 03's execution hardware (tactile fingertips with 3g force detection, palm camera, 20kg payload) and System 0's whole-body control form a strong execution layer; the 200-hour package-sorting marathon proves the execution layer's long-duration stability. This is the layer with the strongest public evidence from Figure.

5.3. L3 Orchestration and Control Layer

40-hour-scale fully autonomous sorting means task-level planning and interruption recovery are inherently handled by Helix; multi-stage household chores (10~15 minutes) in home scenarios are likewise orchestrated directly by the model. Figure has not disclosed an explicit task-orchestration framework, contrasting with Agibot's Genie Studio explicit-workflow approach.

5.4. L4 Memory and State Layer

The continuous operation of long-duration tasks (8→40→200 hours) implies state-management capability; Go-Big's 100,000+ residential-unit first-person video is a cross-scenario "experience library"-style data asset. However, public disclosure of runtime long-term memory mechanisms is missing, marked [To be filled].

5.5. Evaluation and Observation Layer

The BotQ deployment fleet itself is evaluation infrastructure: the deployment count exceeding the employee count means large-scale real-world conditions are continuously producing observation metrics (sorting cadence, failure rates, placement accuracy >99%). Figure has not published a standardized evaluation benchmark or a simulation-evaluation pipeline; evaluation evidence exists in the form of "operational KPIs" rather than "benchmark scores".

5.6. L6 Governance and Safety Layer

RaaS hourly billing ($25 / hour figure is Credibility Level C) is material for commercial governance; but the core public evidence on runtime safety governance is negative: the wrongful-termination lawsuit filed in 2025-11 by a former safety engineer alleges the robot poses a skull-fracture risk, and the case is in the early evidence-discovery stage. This is the group's only litigation-level safety material, indicating that safety governance for industrial humanoids has not yet formed a public engineering standard. Public information on other governance dimensions (permissions, auditing, data compliance) is missing.

5.7. Six-Layer Strength Summary

6. Actual Cases

6.1. BMW Spartanburg Plant

Figure 02 completed an 11-month pilot in the body shop of BMW's Spartanburg plant: supporting production of 30,000+ X3 vehicles, loading 90,000+ sheet-metal parts, about 1,250 running hours, and placement accuracy >99% (per Venture Atlas, 2026-08). The timing of Figure 03's return to BMW for logistics sequencing has three conflicting versions, presented in parallel here without choosing one:

6.2. Catalyst Brands Retail Logistics

On 2026-05-26, Figure signed a commercial agreement with Catalyst Brands (parent of JCPenney, including Aéropostale / Brooks Brothers) to deploy a fleet of Figure 03 units at its Reno, Nevada distribution center, performing retail-logistics sorting together with the Joey Pouch sorting system. As of 2026-08 the two parties have not disclosed the number of units deployed, marked [To be filled]. Together with the Helix 200-hour package-sorting marathon, this case forms capability evidence for the retail-logistics scenario.

6.3. Project Go-Big and Home Scenarios

In partnership with Brookfield, first-person human video is collected across its 100,000+ residential units to pretrain Helix, and zero-shot language-conditioned home navigation (e.g., a "go to the fridge" instruction) has already been achieved. This is the group's only data-collection-type commercial collaboration explicitly designed around "data-for-deployment", pointing to Figure's long-term goal: a general-purpose home robot.

7. Summary

7.1. Strengths and Limitations

Strengths: the capital depth that comes with a $39B valuation; the data flywheel of the BotQ in-house factory plus deployments exceeding the employee count; dual commercialization validation across two scenarios (industrial + retail logistics) with BMW and Catalyst; and Helix's long-horizon operation capability, the strongest in this group (up to 200 hours).

Limitations: the scissors gap between valuation and shipments — a $39B valuation corresponds to cumulative production that has just surpassed 1,000 units; BMW's key commercial terms depend on a Credibility Level C source; the safety lawsuit exposes governance weaknesses; the model is closed-source with no verifiable technical benchmark; and RaaS pricing is opaque.

7.2. Scope of Applicability

Best suited to logistics and manufacturing customers in the North American market that are willing to adopt the RaaS pay-per-hour model and whose scenarios closely match package sorting / material sequencing. For customers with limited budgets who need localized delivery and pricing transparency, the complete-machine sales model of Chinese vendors (Unitree / Agibot / UBTech) has a lower barrier; teams that only need model licensing should evaluate PI (see 08-physical-intelligence.md).

7.3. Selection Recommendations

1. Evaluate Figure using "per-running-hour cost" rather than the full-machine selling price as the budgeting basis, and ask the vendor for a formal RaaS quote.
2. Obtain auditable operational data from the BMW pilot (cadence, failure rate, MTTR) as the acceptance baseline.
3. The contract should include safety-liability clauses and a downtime plan — the industry (including Figure) currently has no public runtime safety standard.
4. Watch its home-scenario progress: if the Go-Big data strategy pays off, Figure may be the first to open the home market, forming a dual growth curve with its industrial line.

Information-Gap Statement

1. Three conflicting versions of the BMW return date (2026-03 / 2026-04 / 2026-06-25 or 07-01); presented in a table in parallel here, not choosing one.
2. Official RaaS pricing is undisclosed; "$25 / robot running hour" is a third-party wiki figure (Credibility Level C), already marked.
3. Catalyst Brands deployment count is undisclosed by both parties, marked [To be filled].
4. Paper-grade architecture material for Helix and System 0: no official blog URL was found; architecture descriptions rely on third-party profiles.
5. Safety lawsuit is in the early evidence-discovery stage; the allegation is a one-sided statement, and the final facts depend on the judicial outcome.

8. References

1. Figure official website (Figure 03 / Helix product narrative) — Figure AI, 2026. https://www.figure.ai
2. Figure AI — Company Profile, Milestones & Funding — Venture Atlas, 2026-08. https://www.ventureatlas.org/company/figure-ai
3. Figure AI Company Profile (2026): Humanoid at Scale — THE CODEW, 2026-08. https://www.thecodew.com/p/figure-ai-company-profile-2026-humanoid.html
4. Figure AI company history (Figure 03 / BotQ / BMW) — AI Wiki, 2026. https://aiwiki.ai/wiki/figure_03
5. Figure 03 sorting parts at the BMW plant, BotQ producing one unit every 60 minutes — Rongrun Mingtuo, 2026. http://www.rrmt.cn/news/79292
6. Humanoid Robots in 2026 Compared — ValueAdd VC, 2026. https://valueaddvc.com/blog/humanoid-robots-in-2026-figure-apptronik-1x-and-tesla-optimus-compared
7. Helix — vision-language-action model technical blog — Figure AI, 2025 (official URL to be verified, listed as "title + institution + year").
8. BotQ mass production and commercialization progress announcement — Figure AI, 2026 (official first-party press-release URL to be verified).

Incremental Update (2026-09-12): Figure AI announced a compute agreement at the scale of 100,000 GPUs (2026-09, per official channels [To be verified]) to support scaled training of the Helix family of VLA models — robotics companies building or locking in large-scale compute capacity is becoming a new variable in "brain-layer" competition, consistent with the "compute-intensive VLA training" judgment in the Chapter 4 architecture of this article.