Robotics Foundation Model Competition Heats Up

로보틱스 | Thu Jul 09 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 4 sources

General Intuition unveiled an embodied AI foundation model trained on gaming data, while Mistral released a navigation model powered by a single RGB camera.

Analysis

[General Intuition] developed an embodied AI foundation model and raised $320M ^[1]^[2]^[3]

Closed a $320M round at a $2.3B valuation
Participation from Bezos
Coatue
Eric Schmidt
Vinod Khosla
and others
Trained on millions of hours of video game play data
Controller input action data is key to spatial-temporal reasoning

[General Intuition] demonstrated zero-shot navigation on a quadruped robot ^[1]^[3]

Fine-tuned with just 8 minutes of real-world robotics data
Uses only a front-facing camera with no additional sensors
Operates in office environments with dynamic obstacles
Aims to serve as a base model for other robotics companies

[Mistral] released the Robostral Navigate robot navigation model ^[4]

8B parameters
self-trained on simulation data
Uses only a single RGB camera
no LiDAR or depth sensors required
Achieved 76.6% success rate on R2R-CE unseen benchmark
9.7 points ahead of single-camera SOTA and 4.5 points ahead of multi-sensor systems

[Robostral Navigate] adopted a pointing-based navigation approach ^[4]

Predicts target coordinates within the camera view
Robust to changes in camera intrinsics and world scale
Generalizes across various embodiments including wheeled
legged
and flying robots
Token-efficient training via prefix-caching
Continuous improvement through reinforcement learning

Robotics Foundation Model Competition Heats Up

Analysis

Sources