AMD released a research paper on September 29, 2026, detailing how reinforcement learning can be applied to train robotic arms using simulation-to-real transfer with Instinct GPUs. It is the company's first significant research update on sim2real techniques since its 2024 MuJoCo JAX blog post.
AMD reported a 2.5x faster training throughput, measured on a single MI300X GPU, compared to a CPU-based setup. That compares with a 1.8x improvement reported in its earlier 2024 blog post.
The research is built on AMD Instinct GPUs and targets robotic automation applications, particularly in industrial and research settings. Availability of the techniques begins with academic and developer access, initially for robotics researchers and developers.
"We trained a pick-and-lift policy for a UFactory XArm-6 with reinforcement learning on AMD Instinct GPUs, then deployed the same network on the real arm with a classical perception front-end in place of privileged simulator state," said Arkojit Ghosh, lead researcher.
The policy's observation space includes proprioceptive state and environment state, with the latter derived from simulated cube position and orientation.
The announcement follows AMD's 2024 blog post on GPU-accelerated RL with domain randomization on a desktop AMD Radeon GPU using MuJoCo JAX. AMD itself frames the significance as an advancement in making real-world robotic control more reliable through simulation.
AMD did not say how the policy would perform on more complex robotic form factors like humanoid robots, and raised the open question of whether the approach could be scaled to more complex tasks. The source says the team plans to explore adding torque or other proprioceptive signals in future work.
Source: [amd](https://rocm.blogs.amd. com/artificial-intelligence/sim2real-rl-instinct/README.html)
