Research @genrobotics_ai || RSS Pioneer 2026 || previously PhD at CMU, @msftresearch, USC, and @CompScienceCU
Seattle, WA
Joined February 2023
- Tweets167
- Following619
- Followers491
- Likes459
Pinned Tweet
Excited to share that I’ve been selected as a member for the Robotics Science and Systems (RSS) Pioneers 2026!
Many academic researchers whose work has influenced me over the years are an alumni of RSS Pioneers, so being part of this cohort means a lot.
See y’all in sydney!
A little update on what I’ve been up to over the past few months!
After finishing my PhD at @CarnegieMellon , I joined @genrobotics_ai , where I’ve been researching robot learning. Building on my PhD research on the safety and reliability of learning-enabled robotic systems, I’m now focused on making robot learning more scalable, robust, and generalizable across tasks and robot embodiments
My work spans the robot learning stack, from scalable data collection systems to real-to-sim pipelines, simulation-based training and post-training for robust autonomy.
Excited to be working with an incredible team on these challenges!
Sad I can’t attend COLM 2026 but if you’re attending, come check out our work led by @Jayzou3773 !
[COLM 2026 AI Measurement Science workshop] Towards Context-Aware Cloud Anomaly Detection: Benchmark, Agent, and Analysis
A CPU spike: an attack or a scheduled backup?
Compared with prior benchmarks, CloudAnoBench exhibits higher ambiguity and greater difficulty, on which both prior machine learning methods and vanilla LLM prompting perform poorly.
Project Page: jayzou3773.github.io/project…
Huge thanks to @Xuan04263449 @HaozeHe_Hector @Parvkpr and other collaborators!
Parv Kapoor retweeted
In our Nature paper, we introduce the first superhuman Stratego AI, which we built using general techniques that we developed for RL & test-time compute under imperfect information. 1/N
from steel city to vice city?
Today, Carnegie Mellon University and Ken Griffin, founder and CEO of Citadel and founder of Griffin Catalyst, announced a historic $3 billion gift to elevate Carnegie Mellon in Pittsburgh and launch Carnegie Mellon University Miami. #CMUMiami
Learn more: cmu.edu/news/stories/archive…
Excited to announce our work on using embedding based temporal safety monitors was accepted to #NeurIPS2026
do they have any paris conference registration tickets left? 🫶🏻 (or i can hold it down in atlanta)
Announcement post coming soon but big thanks to my collaborators! Andy Hammer, Karen Leung, @eskang09, @akapoor_av8r!
Big fan of inference-time enforcement! I still think we’ve underexplored how much VLA behavior can be shaped without retraining. In our ICML 26 work on constrained decoding for nav vlas, we found that even after modifying the action distribution at inference time to satisfy temporally extended constraints, the policy often preserved redundant actions corresponding to alternative (roughly homotopic) ways of completing the task.
Our bottleneck was that the constraints were state-based (but they came with formal guarantees!). This work makes me wonder how far the same idea could go with semantic constraints extracted directly from VLA representations.
Can we steer a robot foundation model without retraining it?
Our new paper explores this question through control-inspired notions of feature observability and controllability in Vision-Language-Action (VLA) models.
We find that state- and action-relevant information can often be linearly predicted from VLA internal representations—and that these representations can be steered at inference time to change robot behavior, without fine-tuning.
As an example, on a real robot, our approach increased the preferred handle-grasp rate from 14% to 74%, while adding only ~1% inference-time overhead.
The broader idea is that representation-level control could provide a lightweight interface for adapting robot foundation models to new preferences and constraints—without retraining.
Joint work with Hugo Buurmeijer Carmen Amo Alonso @SwannAiden
Paper: arxiv.org/abs/2603.05487
#PhysicalAI #Robotics #VLA #RobotLearning
@StanfordEng @StanfordAILab
claude code’s creator using TLA+ was NOT on my 2026 bingo, maybe my time learning TLA+ was not such a waste after all. However, Leslie Lamport (creator of TLA+) in his seminal lectures mentioned writing TLA+ code is more about the process than the artefact and how it encourages systems thinking. Seems like we have outsourced that to claude now.
I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and race conditions. Video attached.
TLA+ also works well. I sometimes combine Lean and TLA+ to look for issues around data flow, concurrency, and state mgmt.
I don't know either language well, but Claude is excellent at both. This approach is super useful for formally modeling your code and finding bugs that a human probably wouldn't have spotted.
Is formal verification the future of coding (or at least, bug finding)?
Fully agree with this take! Given the data gap, the compounding knowledge aspect gathered from multiple deployments is even more important. If agentic systems leave behind reusable models of morphology, interfaces, failure modes, and evaluators, then each deployment can make the next one easier.
Here’s our findings from building robotic harnesses: generalrobotics.company/post…
unpopular opinion but i tried to write a blogpost for one of my ideas and found academic papers to still be the best medium. It was substantially easier for communicating the essence of my idea at a level that felt satisfactory and self contained.
Everyone in academia is feeling the burnout induced by the amount of sloppy and poorly written papers. Surprisingly, llms still arent great at writing/explaining because readability is just not something they optimize for by default. Imo, it’s all the more reason to really spend the time crafting your artefact into a version you’re proud of because all the respectable researchers can now sniff out sloppy writing very quickly.
Seeing a lot of cool applications of GPT 6 Astra so i gave it a go and it kinda blew my mind.
There has been a lot of amazing research recently for real2sim. Going from a single image/video to an interactive simulation environment unlocks evaluation and training capabilities easily (SimFoundry from NVIDIA is a great example, I was waiting for months for the code to be released lol).
I asked Astra to create an environment from a single image to see how good it would do 0 shot and i am 🤯. It even got the color of the liquid so close (although it couldn’t simulate the reaction).
There used to be whole research papers dedicated to getting remotely close to these abilities. Astra can do this 0 shot. Truly crazy times we live in.
check out our write up on how we built this and other cool use cases of robotics harnesses using GRID! generalrobotics.company/post…
thank you for highlighting our work! I believe a pretrained robot policy is implicitly encoding a lot of nuanced motions elicited from demonstration datasets but none of this knowledge can really be explicitly identified. From our harness graphs, it became increasingly clear that the GPT-esque models are building a knowledge base that they leverage each time they see an adjacent task. The nuances are just more interpretable and explicit here.
There’s been a lot of talk lately about whether robotics’ GPT moment could be GPT itself.
Astra can already use tools to control robots, build simulations and write control code. That still seems a long way from a GPT moment, but AI is already taking on parts of a robotics engineer’s job.
On @genrobotics_ai’s GRID, agents determined that a stirring task needed human demonstrations, requested the data and fine-tuned a policy. When perception couldn’t keep up with a moving beaker during pouring, they fine-tuned and integrated a faster tracker.
Those integrations, models and tested fixes stay available for the next task. The team reports about four hours to get the first skill working on a fresh Flexiv setup, with later skills on that same setup deployable in as little as 10–15 minutes.
So I don’t think a stronger GPT automatically replaces embodied AI models. The teams building them can use it to train, test and deploy faster. That could help bring the breakthrough closer.
Parv Kapoor retweeted
For swirling a flask, we gave GRID one phone video of a person doing it. No robot demonstrations.
GRID reconstructed the motion in simulation to ground it in verifiable physics and generated synthetic demonstrations to train a policy.
6/n
Probably a good time to share what I’ve been working on since finishing my PhD at CMU!
When I started my PhD, getting robots to reliably execute complex, long-horizon tasks required enormous amounts of bespoke engineering. Our work on robotics harnesses is accelerating these workflows by an order of magnitude.
One of my favorite examples working on this was our Vid2Sim harness. It takes a single video demonstration and turns it into a policy that learns to reproduce a human swirling motion on the flexiv rizon arms. Swirling is a hard to specify motion and equally hard to teleoperate without special devices.
Check out the blog (and peep my cameos in the video👀)
What if a robot could take a goal and determine how to achieve it?
Today, we’re introducing Auto-Engineering on GRID.
Robotics has more models, simulators, and hardware than ever. Turning them into working deployments still takes specialized engineering.
For two years, we’ve been building that robotics know-how into unified intelligence GRID - spanning robots, skills, models, and proven approaches.
Auto-Engineering puts it to work by completely automating how robot intelligence is built, tested, and deployed.