◆ Gentoo Logic · Modeling Warehouse

Architecture comparison

Figure vs Physical Intelligence

Figure runs closer to the foundational model (Native (own / self-hosted weights)) than Physical Intelligence (Native (own / self-hosted weights)). Shared foundation: proprietary / self-built models (highlighted).

Figure

Robotics

Proximity
Native (own / self-hosted weights)
Models
proprietary / self-built models
Cloud · datastore
Hybridproprietary manufacturing/operations data infrastructure (MES/PLM/ERP/WMS stack)[2]custom telemetry and training data pipelines for Helix VLA (vision, language, proprioception)[7][9]
Compliance
Not publicly specified; likely standard enterprise practices but no explicit SOC2/ISO27001/IEC 62443 claims in accessible materials[1][2][7][9]
Architecture

Figure’s production architecture is centered on its in-house **Helix** vision-language-action stack running on vertically integrated humanoid hardware and a proprietary BotQ manufacturing/software infrastructure, with robots streaming high-bandwidth data for continuous training and fleet improvement[1][2][7][9]. Core control and reasoning have moved off external LLMs to fully in-house Helix models, with cloud services used as supporting infrastructure rather than as foundational AI engines[7][9].

Physical Intelligence

Robotics

Proximity
Native (own / self-hosted weights)
Models
proprietary / self-built models
Cloud · datastore
HybridObject storage (WEKA Data Platform over high-performance files/object)[2]Oracle Cloud Infrastructure storage/DB services (unspecified mix, likely OCI Object Storage and managed DB)[1]Custom distributed data infrastructure for robot learning (data pipelines between raw telemetry and training/eval)[11]
Compliance
Architecture

Physical Intelligence runs a proprietary robotics foundation model stack with large-scale training and data infrastructure, using WEKA for high-performance data and Oracle Cloud Infrastructure for hosted compute, complemented by Kubernetes-based CI/CD and GPU-heavy cloud environments.[1][2][7][11][12] The architecture centers on Python/C++ robotics services, PyTorch/JAX-based model training on cloud GPUs, and data pipelines from robot teleoperation and simulation into foundation model training and evaluation.[11][12]

← Full orbit map · Score your own stack →