Robotics Foundation Models & Humanoids — Market

Verified claims and product-axis read for Robotics Foundation Models & Humanoids. Every fact below is sourced; every product judgment traces back to underlying signals.

Verified facts

• Apptronik unveiled a wheeled industrial mobile manipulator called 'Astra' in addition to Apollo in 2024. ↗ _(historical_event)_
• DeepMind's RT-X dataset combined 22 robot datasets and trained a single policy that transferred zero-shot. ↗ (other)
• AgiBot () announced production of 1,000 humanoid units by end of 2024. ↗ _(historical_event)_
• Tesla used Optimus to serve drinks at the 'We, Robot' event in October 2024 with humans teleoperating, not autonomous control. ↗ (other)
• AgiBot's founder Peng Zhihui ('') was previously a Huawei Genius Youth program member. ↗ (other)
• 1X Technologies is headquartered in Moss, Norway, with US operations in Sunnyvale. ↗ (other)
• Apptronik's Nick Paine and Jeff Cardenas founded the company in 2016, ~8 years before the humanoid wave. ↗ _(historical_event)_
• Figure AI moved its HQ to a larger Sunnyvale facility in 2024 to support volume manufacturing prep. ↗ _(historical_event)_
• Skild AI partners with multiple humanoid OEMs but does not name them publicly, per founder interviews. ↗ (other)
• Physical Intelligence raised a $400M round in late 2024 without owning any robot hardware. ↗ (financial)

Top products (engine read)

Isaac Robotics — simulation + training workflow stack for robot foundation models

Opportunity: Sim-to-real data is the #1 bottleneck cited in cbr_post e9295e2a; Unitree explicitly opened a 'Senior Workflow Specialist - Isaac Robotics' role and Apptronik is staffing simulation, meaning OEMs are standardizing on NVIDIA's sim pipeline for foundation-model training rather than rolling their own.

Isaac is the shared training/eval substrate underneath the VLA race; the A1 evidence that even Chinese (Unitree) and US (Apptronik) humanoid teams are hiring around it confirms it as the default tooling layer.

Helix — humanoid VLA foundation model + whole-body control stack

Opportunity: Figure is staffing a dedicated Helix training-infra org plus whole-body-control infra hires while publicly committing to 100k humanoids in 4 years and a BMW Spartanburg deployment; the open question (demand 364513d2) is whether VLA policies generalize from demo skills to real housework / line-side tasks.

Helix is the in-house humanoid foundation model that Figure is betting its $675M round and Sunnyvale scale-up on; the A1 stack shows training-infra + whole-body-control are the bottleneck they are hiring against, not perception alone.

Gemini Robotics — VLA built on Gemini 2.0

Opportunity: Hyperscaler-grade VLA pulls humanoid foundation models into the same release cadence as frontier LLMs; demand_query 4ad70a3c shows the audience is already grouping 'Nvidia foundation model for humanoids' and Gemini-class models into one news bucket.

Gemini Robotics is the cloud-AI lab's wedge into the foundation-model-for-robots race; it competes with Helix/π for which model becomes the default brain that humanoid OEMs ship behind.

NVIDIA Cosmos — foundation model + world-model stack for physical AI

Opportunity: Cosmos pitches a foundation model that runs on workstation GPUs, addressing the explicit gap raised in cbr_post 76342328 that no world-model has booked revenue yet — robotics training/eval is the first plausible market.

Cosmos is NVIDIA's bid to be the Android of physical AI — it pairs naturally with their Isaac sim stack (see Unitree's Isaac job) and undercuts cloud-only VLAs by targeting on-prem RTX PRO inference.

See the Products and Strategy modules for the full product list and forward-looking judgment.