Michelle Lu, co-founder and CEO of Vsim Technology, helped create Isaac Gym, described as the world's first end-to-end GPU-accelerated reinforcement learning framework, while she was Director of Simulation Technology at NVIDIA. On 2 October 2026 she appeared in an interview with host Claire to discuss robot learning simulation: how simulation and artificial intelligence can be used to teach robots new skills.
From NVIDIA to Vsim
Lu spent part of her career at NVIDIA as Director of Simulation Technology. There she and her co-founder built Isaac Gym, an end-to-end reinforcement learning framework that runs on GPUs. The framework set the template for how modern GPU-accelerated, simulation-based reinforcement learning is carried out in robotics.
In 2022 Lu and her co-founder left NVIDIA to start Vsim. The company is now the vehicle for her work on teaching robots through simulation. She serves as its CEO.
The interview was framed around that shift, from building a widely used training framework inside a large chip company to running a startup focused on the same broad problem.
The problems Vsim is aiming at
Lu and her co-founder started Vsim to address what the interview describes as the limitations that plague simulation-based robotics AI. Three are named: efficiency, scalability and the sim-to-real gap.
Efficiency concerns how much computation and time it takes to train a robot policy in simulation. Scalability concerns whether that training can grow to more tasks and more complex robots. The sim-to-real gap is the difference between how a robot behaves in a simulated environment and how it behaves when the learned skill is moved onto physical hardware.
Reinforcement learning is the method at the centre of this work. A robot, or a simulated version of it, tries actions and is rewarded for outcomes, and over many repeated attempts it learns a skill. Running those attempts in simulation avoids the wear, cost and risk of running them on physical machines. That is the reason simulation sits at the centre of the approach Lu describes.
What the interview covers
The conversation centres on using simulation and artificial intelligence to teach robots new skills. The material available about the episode does not give a detailed account of Vsim's technology, products, customers, funding or timelines, and none are reported here.
The episode is number 164 in a weekly series that explores robotics, artificial intelligence and autonomous machines. Later episodes in the series have featured guests working on construction robots, evolving robot bodies and brains, and robots used in scientific facilities, which places the Vsim conversation within a broad run of research and industry topics.
Why it matters, and what remains unproven
Lu's background gives the interview weight on one specific point. Isaac Gym is credited with setting the template for GPU-accelerated reinforcement learning in robotics, so the person who helped define that method is now building a company around its remaining weaknesses. When a framework's creators leave to address its limits, that is a useful signal about where they think the unsolved work lies. In this case the stated list is short: efficiency, scalability and the sim-to-real gap.
The list is also notable for what it leaves out. Training speed was the headline gain of GPU-accelerated simulation, yet efficiency is still named as a limitation. That suggests the first generation of tools did not close the problem, even if it changed the standard way of doing the work. The sim-to-real gap is the more persistent issue for any team that trains in simulation and then deploys on hardware, because a policy that performs well in a virtual environment can fail on a physical robot.
What remains unproven is whether Vsim has solved any of the three problems. The available account states the company's motivation, not its results. There are no published benchmarks, no named deployments and no comparison with other simulation platforms. Without those, the company's claims cannot be compared with existing approaches, and a reader should treat Vsim as a startup with a credible founding team and a clearly stated target, not as a demonstrated solution.
The practical test will be evidence. Useful measures would be training time for a given task, the range of robots and skills that can be trained at scale, and how often a policy trained in simulation works on real hardware without extensive retuning. Until Vsim or its customers report figures of that kind, the interview is best read as a statement of intent from someone with direct experience of how the current tools came about.