Amazon Frontier AI & Robotics

亚马逊前沿 AI 与机器人团队FARAdvanced

Amazon's robot foundation-model team, built around the Covariant team it absorbed.

Amazon FAR is Amazon's internal frontier robotics research team. In August 2024, Amazon licensed the technology of the AI-robotics company Covariant and hired its founders — Pieter Abbeel (a UC Berkeley professor), Peter Chen, and Rocky Duan — along with roughly a quarter of its staff; this group became the core of FAR, which researches robot foundation models with a focus on whole-body control and data generation for humanoid robots. Its results include Holosoma, a framework for training humanoid reinforcement learning in simulation and deploying it to the real robot, and OmniRetarget, a motion-retargeting method. Abbeel reportedly later moved to lead frontier-model research within Amazon's AGI organization. FAR is a separate effort from Amazon's warehouse and logistics robotics business, Amazon Robotics.

ExampleHolosoma packages a humanoid robot's simulation training and real-robot deployment into one open-source pipeline.

Also called: Amazon FAR
Related: Covariant · Amazon Robotics · Holosoma (Amazon FAR humanoid RL framework) · OmniRetarget · Whole-Body Control · Humanoid Robot
Sources: Amazon hires from AI robotics startup Covariant, licenses technology
Pieter Abbeel - Amazon Science
As of: 2026-09