OpenAI is getting back into robotics

Three years ago, OpenAI gave up on physical robots to focus on pure software and language. Now, with ChatGPT sitting at the center of the tech world, the company is finally ready to step back into the physical world.

OpenAI is officially rebooting its robotics research group. As first reported by Fortune, the AI powerhouse has begun hiring again, opening up roles for research robotics engineers.

This marks a significant return to a domain the company walked away from in July 2021, well before the generative AI wave took over.

The missing pieces of AGI

When OpenAI dissolved its original robotics team three years ago, the timing seemed puzzling to outsiders. But the team hit a wall.

At the time, co-founder Wojciech Zaremba explained that while he believed in the robotics approach, key pieces were missing from an AGI standpoint. The team had hoped to get far using self-generated data and reinforcement learning, but the foundation simply was not there yet.

Fast forward to today, and large multimodal models have changed the game completely. The missing components Zaremba talked about look a lot like the vision and reasoning capabilities OpenAI spent the last three years building.

Building minds, not metal

Do not expect an OpenAI-branded mechanical arm or bipedal robot anytime soon. The company is sticking to what it does best: building the underlying intelligence.

According to the job listing, OpenAI is looking for engineers to train multimodal robotics models to unlock new capabilities for robots built by external partners. The role involves designing new model architectures, gathering robotics data, and running evaluations.

OpenAI stated that candidates need a strong research background paired with experience shipping actual AI products.

This move connects the dots on earlier signals. In March 2023, OpenAI launched APIs that let developers connect physical robots with generative AI models. More tellingly, OpenAI participated as an investor in humanoid startup Figure AI during its Series B round earlier this year.

Why it matters

We think this is the most logical next frontier for OpenAI. True general intelligence cannot stay trapped behind a chat prompt forever; it needs to perceive, interact with, and manipulate the physical world.

For builders and end users, this does not mean buying an OpenAI robot tomorrow. Instead, it means companies like Figure and other hardware partners will soon have access to cutting-edge multimodal backbones designed specifically for spatial reasoning and tool use.

Across the industry, this puts immense pressure on rivals. Google DeepMind has pushed hard into robotics, and countless hardware startups are racing to build humanoids. OpenAI entering the fray validates embodied AI as the defining battleground of the next decade.

Strategy-wise, OpenAI is playing it smart by avoiding the hardware trap. Manufacturing mechanics and managing supply chains is notoriously capital-intensive. By letting partners handle the hardware while OpenAI supplies the brains, the company gets access to valuable physical-world data without taking on manufacturing risk.

That said, gathering high-quality physical data remains a brutal bottleneck. Whether modern multimodal models can overcome the messiness of the real world better than reinforcement learning did three years ago is the real question to watch.