There Is Plenty of Room (and Pain!) in Physical AI

01 — Looking for Pain Points

In 1959, physicist Richard Feynman gave a talk called “There's Plenty of Room at the Bottom” where he described interesting physics problems not being worked on. His intent was not to wow the world with a breakthrough, but point at a wide open field and say "Look at all these problems nobody's working on!"

I love that story because it reminds me that entrepreneurs thirst for problems to solve. The question is: which one?

Sitting here in 2026, as code gets cheaper and cheaper, a longstanding truth stands the test of time: best gift for an entrepreneur is a problem with a sharp pain and real demand. Code, investment, hiring, all are secondary.

Unlike Feynman's research focus, we focused on fellow entrepreneurs and builders, and we compiled an initial list of the top pain points we have seen in the physical AI space. Consider this an open call for problem solvers, a lightning rod for creative people.

02 — How We Created This List

We at Haptic are working on improving the infra for physical AI synthetic data and training, so we are always talking to companies in the physical AI space to understand their pain points.

Among the many resources in Physical AI space, few are as high signal as the RoboPapers podcast by Chris Paxton and Michael Cho. Their work is excellent.

We wondered: what problems, pain points, and surprising take-aways already exist, lurking in the podcast’s deep conversations?

We used Claude Code to extract every time a pain point or problem was mentioned in the 64 episodes of the podcast (and their associated papers). AI was not perfect, so we manually reviewed and cleaned up the list, but it is a reasonable starting point.

Now any curious or entrepreneurial hacker can take a look at this unofficial “request for products”. Yes, the RoboPapers podcast is a research-focused one, so the problems lean towards the research side, but we will be sharing more industrial ones as we navigate the space.

03 — Top 10 Pain Points in the Physical AI frontier

We ranked these by how often each pain point was mentioned across 64 RoboPapers episodes. The number in parentheses shows mentions per episode count (for example, 22/64).

1. Scalable robot (and human-robot) data collection (22/64) - Collecting high-quality robot data is still slow, expensive, and hard to scale.
2. Generalization and zero-shot robustness (12/64) - Policies often fail when objects, tasks, or environments shift beyond training conditions.
3. Dexterous and contact-rich manipulation (10/64) - Multi-finger control and force-aware contact handling remain difficult in real tasks.
4. Teleoperation and whole-body data collection (10/64) - Current teleop setups are uncomfortable, limited, and hard to scale for whole-body behavior.
5. Sim-to-real and simulation environment creation (10/64) - Building useful sims takes major effort, and transfer to real robots is still fragile.
6. Evaluation and benchmarking at scale (9/64) - Reproducible real-world evaluation is costly and hard to standardize across labs.
7. VLAs, foundation models, and world models for control (8/64) - General-purpose models still struggle with reliability, 3D reasoning, and control alignment.
8. Human video / human-to-robot transfer (6/64) - Human demonstrations lack robot-ready actions, dynamics, and embodiment compatibility.
9. Long-horizon and memory (6/64) - Most policies are weak on long sequences and memory-dependent decision making.
10. RL scaling and offline-to-online (6/64) - Exploration, data efficiency, and pushing reliability toward deployment-grade performance remain open.

04 — But Why Share This Publicly?

As we showed friends this blog post, this was a common question and our answer was simple:

1. We do not own this data. It is already out there.
2. We believe in open source physical AI... so holding back any learnings would be antithetical.
3. If someone finds the same pain point, builds a better solution, and out-executes us, that is on us. The ecosystem wins.

On the flip side, we are new to the physical AI ecosystem, and we believe strongly in a “leave things better than you found them” mentality, so we want to start to improve things.

05 — Surprising Takeaways

A few things stood out once we aggregated all 64 episodes:

1. A common bottleneck is not “ideas” or even “model architecture”, just boring data plumbing. We have heard this a lot... but we are constantly surprised.

2. Sim-to-real is still painful enough that many teams pay the real-world data tax. We believe in the power of simulations and synthetic data, so we were surprised how painful the current tech still is. Or to put it another way, a researcher said to us: “today synthetic data is probably good enough to publish, but not good enough to deploy.”

06 — What's Next

We will be updating this repo, publishing more learnings and problems, and sharing some of the projects we are working on.

Stay tuned!

07 — We Want to Hear From You

We would love to hear from you if...

1. You agree/disagree strongly with any of these problems... save entrepreneurs from wasting time! Even social media threads about why “the haptic guys don't know what is what” can help everyone learn.
2. You are interested in fixing any of these… or already do! We can give you some shout outs.
3. You know more important or painful problems.

Passion in the domain is more important than agreeing with us.

Reach out via email, X, or GitHub.

Special thanks to Chris Paxton, Craig Cheney, and Anthony Avedissian for reviewing this blog post and giving us feedback.