Joining Physical Intelligence to Work on Simulation

What makes researching simulation at Pi exciting!

In a surprising turn of events, I am now not only completing my short PhD in a matter of months, I am also joining Physical Intelligence (Pi) full time after a short internship starting now. I’ll leave my short PhD as a story for another time.

Until recently, I had reservations about Pi. From the outside, it appeared that the company was not particularly interested in simulation. That could not be further from the truth. Pi offers an unusually exciting environment for simulation research, and its culture and ambitions ultimately convinced me that it was the right fit. (And we are hiring!)

There are many typical reasons why a research company might be a great place to work, including an open research and communication culture and a high density of talented people. Pi has many of these, along with several qualities that make it distinctly unique to me.

Research Trend Setting

One quality that makes Pi unique is its ability to shape research trends. The influence of many of Pi’s researchers is arguably one reason robotics research today places greater emphasis on VLA models and teleoperation than on simulation scaling or sim-to-real manipulation.

Regardless of where you stand on the right way to “solve” general robotics, the ability to influence research trends creates a favorable ecosystem for Pi to accelerate research. There is a higher likelihood of new team members being aligned with our vision for generalist robots and having the relevant knowledge to work at Pi. There is a higher likelihood that Pi can benefit from the open-source research conducted elsewhere. There is also a higher likelihood that our models can integrate more easily with other companies, projects, workflows etc. because the surrounding infrastructure increasingly supports similar robotics approaches as Pi.

Now as a representative of simulation research at Pi, I hope I can help steer simulation research towards directions with higher signal. I certainly have plenty of hot-takes about simulation evaluations in the age of robotics foundation models.

Simulation at Pi

To perhaps the surprise of many, I believe that Pi deeply understands the value of simulation and the breadth of possible use-cases.

I actually spent most of my first 2 weeks at Pi not writing any useful code. The first week was filled with conversations with researchers and engineers, each with different expectations and requests for simulation. The following week consisted of staring at 1000s of real evaluation videos and metrics, mapping out how simulation can both complement and inform our real evaluations. Those two weeks made it clear that there is an enormous range of valuable simulation infrastructure and research problems worth exploring at Pi.

The breadth of possible use-cases of simulation underpins the benefit of investing research/engineering hours into scaling up simulation to accelerate robotics research. Possible directions include task generation, simulation-data co-training, simulation-based scaling laws, and so many others. As an anecdote, I often tell people my job at Pi is to ensure everyone else can do their research faster.

Ownership and Autonomy

Even in my short time at Pi, I have felt a strong culture that encourages researchers to take ownership in their research directions. People here are never waiting for feedback from an “advisor” and always find ways to be unblocked or to parallelize their work. Each researcher here will autonomously decide what is the best angle of attack to improve anything, from machine learning to even user interfaces.

Pi is also generally quite careful with how it builds the team. One benefit beyond maintaining a high talent density is that Pi can place a lot of faith in even the new members to make the right bets and decisions in research. This level of ownership and trust really surprised me. On my first day, I was told that that I could choose to rewrite large parts of our simulation infrastructure if I felt I could improve it.

As I near the end of my PhD, I have realized how much I value autonomy in my work. I believe the culture of ownership and autonomy at Pi provides the right environment to ensure we continue to do great research. Importantly, Pi continues to look for great people who fit that culture, and I hope that continues as we tackle the hardest challenges in robotics.

Seeking Novel Challenges

Throughout my life, one of the objectives I have implicitly optimized for was looking for challenges. Without something difficult to work toward, I quickly become bored and unmotivated.

The first example is a shift from working on full-stack engineering (frontend/backend, databases etc.) to working on AI. As full-stack engineering became more familiar, the less interest and time I spent on making the fun apps that got me hooked into programming in the first place.

The next example was starting a PhD instead of going to industry out of undergraduate. I had opportunities to work in machine learning full time, but I was drawn to the novel challenges of research. Not knowing whether a solution existed, or whether an entire direction would succeed, created the kind of risk and reward that I have come to enjoy.

Fast forward to today, joining Pi is another massive and new challenge for me. I could have joined a place where I was one of many researchers in simulation. Instead I chose to work at Pi where I was one of few. By being one of the few people working on simulation at Pi, there’s extra responsibility given to me on so many different things. Whether it’s Pi’s long-term sim strategy or figuring out the right way to build a scalable sim framework, there is always something new and challenging to think about. That responsibility has become another source of motivation every day. To date there has not been a single dull day at Pi.

---

Ultimately, Pi’s research influence and culture, along with unusually rich opportunities for challenging simulation infra/research made Pi a uniquely compelling place for me.