What tech stack does Physical Intelligence use?

Physical Intelligence's primary ML framework is JAX — the functional deep learning library developed by Google, which enables automatic differentiation and JIT compilation via the XLA compiler. JAX was the original framework for the openpi repository, reflecting the Google DeepMind heritage of the founding team. In September 2025, Physical Intelligence added full PyTorch support to openpi, making its model suite accessible to the broader research community that predominantly uses PyTorch for prototyping and fine-tuning workflows.

For compute infrastructure, Physical Intelligence trains multi-billion-parameter VLA models on high-performance GPU clusters — H100 and H200 class hardware — for both supervised pre-training and online reinforcement learning (introduced with π*0.6 in November 2025). The company's JAX dependency and the founding team's DeepMind origins strongly suggest significant Google Cloud Platform usage, likely including TPU pods for large-scale pre-training runs. Job descriptions have explicitly called for experience managing GPU/TPU compute, building and optimizing JAX training pipelines, and debugging distributed training failures across hundreds of accelerators.

For robotics middleware, ROS (Robot Operating System) is the de facto standard at the hardware-software interface layer across the industry, and Physical Intelligence's openpi repository explicitly supports robot platforms that use ROS-based drivers and control stacks — including Franka Panda arms (via the DROID dataset), ALOHA dual-arm platforms, and mobile manipulators like the Stanford Mobile Aloha. The π0 model architecture itself is a 3-billion-parameter base vision-language model coupled with a flow matching diffusion action head — an end-to-end trainable system that processes multi-camera RGB observations and natural language instructions to produce continuous joint-space action trajectories.

For training data, Physical Intelligence uses a combination of the Open X Embodiment Dataset (a large open-source collection of robot manipulation episodes in RLDS format), internet-scale vision-language pretraining data from public sources, and a proprietary π Dataset collected in-house across at least eight different robotic platforms in their San Francisco lab. The openpi repository documents fine-tuning recipes on the DROID dataset (Franka arm, diverse real-world environments), the ALOHA platform (dual-arm, dexterous tasks), and the Libero benchmark (simulation). By November 2025, the training corpus across π0.5 included data from seven platforms, 68 tasks, and 104 real-world homes — a diversity that defines the cross-embodiment learning claim.

For model distribution, Physical Intelligence uses Hugging Face Hub for publishing model weights (π0 and π0.5 base weights are publicly downloadable) and GitHub for code (Physical-Intelligence/openpi has accumulated significant stars and forks from the research community). The FAST action tokenization method — released alongside π0 open-sourcing — is published as a standalone paper and available in openpi, providing a 5x speedup on autoregressive VLA training that third parties can adopt independently.

On the commercial and GTM tooling side, Physical Intelligence is a pre-sales-team company. No CRM, sales engagement platform, or marketing automation stack has been publicly detected. With COO Lachy Groom managing partnerships and business development directly, the company almost certainly uses standard productivity tools (Google Workspace, likely Notion or Linear for internal tracking). As the commercial team builds out, adoption of Salesforce or HubSpot and a sales engagement platform would be the natural progression — and would constitute a procurement opportunity for relevant vendors.

For ML infrastructure and cloud vendors, the JAX-plus-GPU/TPU profile is a direct buy signal. Physical Intelligence is a compute-intensive customer with sophisticated ML engineering requirements and genuine budget (over $1 billion raised). The highest-priority displacement plays are cloud GPU infrastructure (H100/H200 clusters on GCP or multi-cloud), TPU pod access for JAX-native large-scale training, and MLOps tooling that supports distributed JAX experiments. Pitches should lead with performance per dollar on large VLA training runs, JAX-native optimization, and support for online RL workloads — the latter being a newer, less-commoditized requirement that emerged with π*0.6.

For data and annotation vendors, Physical Intelligence is actively collecting proprietary dexterous manipulation data across multiple robot platforms in its SF lab and through commercial partner deployments. Data quality, teleoperation infrastructure (the company uses teleop for human demonstration collection), episode storage and replay tooling, and annotation pipelines for vision-language alignment are all active procurement categories. The research team's academic background means they will evaluate tooling rigorously — provide a research-grade option alongside any commercial offering.

For enterprise software vendors (security, identity, productivity, CRM): the company is still research-culture-first with a founders-run procurement process. Pure enterprise software that cannot be self-served or trialed by a technical user will face a high friction barrier. The best entry points are warm introductions through shared investors or through the open-source research community that actively uses openpi. Monitor for VP-level hires in sales, customer success, and business development as the signal that the enterprise software procurement process is formalizing.