🎯 Move over grid-resolved Hamilton–Jacobi (HJ) backward reachable sets/tubes (BRS/BRTs): With a Cole-Hopf-type transformation of the viscous HJ IVP, turns out Feynman-Kac's formula may recover BRTs as a quasilinearized Gaussian-kernel density representation. In a Picard iterative scheme, we find that 🛡️☠️➡️✨ combinatorial grid explosion reduces to linear Monte-Carlo sampling. Results?😏 ✅ 📊 dimension-independent convergence • 💾 worry-free, memoryless optimization compared to levelsets classical compute baselines • 🔓 tight L²/Hausdorff error bounds from levelset methods• 🤖 and efficacy on benchmarks previously considered computationally intractable 👉➡️ up to 100,000 European starlings (sturnus vulgaris) in multi-predator (peregrine falcon) games --- on commodity multicore CPUs --- inspired by recorded murmurations over Rome, Italy. Solves for phase transitions under vacuole nucleation, cordon formation detection, and flocks splitting🌍. 🧠⚡ TL;DR: we may now teach fleets of autonomous robots/cars not just what to do, but where danger geometrically lives. 🌌🏭 HJ-Gauss opens the door to scalable, certifiably trustworthy embodied AI safety verification. 🧑💻 The full codebase is live and open-source at github.com/robotsorcerer/levelsetpy --- clone it and certify your own swarms! Scalability über alles!🎩👌🤯
🎯 Move over grid-resolved Hamilton–Jacobi (HJ) backward reachable sets/tubes (BRS/BRTs): With a Cole-Hopf-type transformation of the viscous HJ IVP, turns out Feynman-Kac's formula may recover BRTs as a quasilinearized Gaussian-kernel density representation. In a Picard iterative scheme, we find that 🛡️☠️➡️✨ combinatorial grid explosion reduces to linear Monte-Carlo sampling. Results?😏 ✅ 📊 dimension-independent convergence • 💾 worry-free, memoryless optimization compared to levelsets classical compute baselines • 🔓 tight L²/Hausdorff error bounds from levelset methods• 🤖 and efficacy on benchmarks previously considered computationally intractable 👉➡️ up to 100,000 European starlings (sturnus vulgaris) in multi-predator (peregrine falcon) games --- on commodity multicore CPUs --- inspired by recorded murmurations over Rome, Italy. Solves for phase transitions under vacuole nucleation, cordon formation detection, and flocks splitting🌍. 🧠⚡ TL;DR: we may now teach fleets of autonomous robots/cars not just what to do, but where danger geometrically lives. 🌌🏭 HJ-Gauss opens the door to scalable, certifiably trustworthy embodied AI safety verification. 🧑💻 The full codebase is live and open-source at github.com/robotsorcerer/levelsetpy --- clone it and certify your own swarms! Scalability über alles!🎩👌🤯
For long-horizon planning tasks, finite-dimensional diffusion policies wack. Well, this infinite-dimensional policy smacks! 🏭 My 👉 recent work moves beyond point-based sampling/estimates in diffusion to function-space diffusion + formal control verification, delivering a principled framework for high-stakes robots-enabled production control systems. Results? 🎯 Improved convergence in robot manipulation tasks. There's more: in supply chain manufacturing, certifiable bottleneck predictions and the reliable dispatch optimization of utilization efficacy also leveled up: CONWIP with certified safety under prediction uncertainty 💪 • 🔍 🚀Bottleneck detection with theoretical grounding • 📈 Smooth dispatch policies ✅ for resilient, efficient supply chains.🏆 Some good Samaritan's AI agent made a neat summary here.🎤⬇️
For long-horizon planning tasks, finite-dimensional diffusion policies wack. Well, this infinite-dimensional policy smacks! 🏭 My 👉 recent work moves beyond point-based sampling/estimates in diffusion to function-space diffusion + formal control verification, delivering a principled framework for high-stakes robots-enabled production control systems. Results? 🎯 Improved convergence in robot manipulation tasks. There's more: in supply chain manufacturing, certifiable bottleneck predictions and the reliable dispatch optimization of utilization efficacy also leveled up: CONWIP with certified safety under prediction uncertainty 💪 • 🔍 🚀Bottleneck detection with theoretical grounding • 📈 Smooth dispatch policies ✅ for resilient, efficient supply chains.🏆 Some good Samaritan's AI agent made a neat summary here.🎤⬇️
New on the blog ✍️📈 I wrote up the ideas behind RoboDiff as a full deep-dive: Infinite-Dimensional Diffusion Policies via Inverse Kolmogorov PDEs. Why point-based diffusion policies drift on long horizons, how lifting the whole diffusion to function space fixes it, the backward Kolmogorov equation the denoiser secretly learns, the three-line Cameron-Martin code change — and the results (17% higher PushT reward, 67.6% lower drift, 96% fewer deadlocks). 🧮🤖
New on the blog ✍️📈 I wrote up the ideas behind RoboDiff as a full deep-dive: Infinite-Dimensional Diffusion Policies via Inverse Kolmogorov PDEs. Why point-based diffusion policies drift on long horizons, how lifting the whole diffusion to function space fixes it, the backward Kolmogorov equation the denoiser secretly learns, the three-line Cameron-Martin code change — and the results (17% higher PushT reward, 67.6% lower drift, 96% fewer deadlocks). 🧮🤖
Full circle of life moment 🔄✨ I’m back in Amazon Robotics! 🤯🧠⚙️ #RobotsRockAndRoll.
🚨 Game theory has officially met ER practice 🏥 + 🎮 = ❤️🚀 What if trauma resuscitation teams could coordinate efforts better like a well-oiled machine? Perhaps, healthcare workers' workflows may become: ✅ fairer (balanced workloads ⚖️) ✅ smarter (skill-task alignment 🧠) and ✅ efficient with (rapid, stable decisions ⚡) 🎯 Our latest work addresses these in a distributed generalized Nash equilibrium (GNE)-seeking mechanism, grounded in real clinical workflows at Weill Cornell Medicine. The result? A principled framework for orchestrating high-stakes team dynamics under time pressure, resource constraints, and diverse expertise — delivering the best possible outcomes in real time. 🏆🌍🚀 Now headed to IFAC World Congress! 🎉
🚨 Game theory has officially met ER practice 🏥 + 🎮 = ❤️🚀 What if trauma resuscitation teams could coordinate efforts better like a well-oiled machine? Perhaps, healthcare workers' workflows may become: ✅ fairer (balanced workloads ⚖️) ✅ smarter (skill-task alignment 🧠) and ✅ efficient with (rapid, stable decisions ⚡) 🎯 Our latest work addresses these in a distributed generalized Nash equilibrium (GNE)-seeking mechanism, grounded in real clinical workflows at Weill Cornell Medicine. The result? A principled framework for orchestrating high-stakes team dynamics under time pressure, resource constraints, and diverse expertise — delivering the best possible outcomes in real time. 🏆🌍🚀 Now headed to IFAC World Congress! 🎉
Spoke at the Chan-Zuckerberg Initiative in San Francisco on Reinforcement and Supervised Learning in Medical Physics & Engineering. Slides.
Delivered a talk at Agility Robotics Salem, OR on Nonlinear, Singularly Perturbed Control for the Humanoid Platform: DIGIT.
Delivered a talk at the Robotics & AI Institute in Cambridge, MA on Towards Ubiquitous Robotics Automation. Slides.
Hearty congratulations to our 2024 intern and resourceful research collaborator Abulikemu Abuduweili. Abu successfully defended his PhD thesis at the CMU Robotics Institute last week. Later in the week, he would sell his soul for a handsome amount to the ultimate devil join Apple as a Robotics Research Scientist/Engineer. Keep shining those bright white lights, Abu!
Hearty congratulations to our 2024 intern and resourceful research collaborator Abulikemu Abuduweili. Abu successfully defended his PhD thesis at the CMU Robotics Institute last week. Later in the week, he would sell his soul for a handsome amount to the ultimate devil join Apple as a Robotics Research Scientist/Engineer. Keep shining those bright white lights, Abu!
Talked at Wayve Technologies in London on System Identification for Planning in Reinforcement Learning.
Presented at Google DeepMind Robotics, San Francisco on State Representation in Reinforcement Learning. Slides.
Slides from my ACC 2025 Denver talk.
Presented at META Reality Labs in San Francisco on Robustness and Efficient State Representation in Open Embodiments. Slides
I am attending the next American Control Conference (ACC) in Denver, Co. Among other things, I will be (1) chairing the regular session of nonlinear systems' stability; (2) a panelist on the ASME industry-oriented student special session for students hoping to accelerate their career into industry; and (3) presenting this recent layered nonlinear control paper. Looking forward to hearing and learning from the leading minds in our community.
I am attending the next American Control Conference (ACC) in Denver, Co. Among other things, I will be (1) chairing the regular session of nonlinear systems' stability; (2) a panelist on the ASME industry-oriented student special session for students hoping to accelerate their career into industry; and (3) presenting this recent layered nonlinear control paper. Looking forward to hearing and learning from the leading minds in our community.
Entertain your mind reading this newly accepted ACM Transactions on Mathematical Software (TOMS) journal article. It describes our GPU-accelerated Levelsetpy safety analyses software package.
Towards fast strain regulation in continuum robots, in this new ACC 2025 paper we argue for and demonstrate the efficacy of a decentralized continuum dynamical system whose whole-body strain dynamics is resolved in an adaptive and hierarchical nonlinear control fashion.
Together with Volker Pohl, I am co-chairing the computational methods in control theory at the next Conference on Decision and Control in Milan (December 2024). Evviva!
Recent talk at McGill's Mechanical Engineering and MILA, on a layered architecture and time-scale separation control scheme in the quest towards the real-time control of heterogeneous soft robotics devices.
Two papers accepted to the IEEE Control and Decision Conference, CDC 2024, Milano, Italia! Evviva!
Will you be at the next ICRA conference? I am but a weave within the rich tapestry of excellent leaders in the soft robotics community, fashioned together to disseminate the latest findings -- new hardware fabrics, mathematical/ML models, and control tools -- in a workshop on biologically-inspired embodied intelligence for modern robots and devices. Please see the workshop webpage for details.
Will you be at the next ICRA conference? I am but a weave within the rich tapestry of excellent leaders in the soft robotics community, fashioned together to disseminate the latest findings -- new hardware fabrics, mathematical/ML models, and control tools -- in a workshop on biologically-inspired embodied intelligence for modern robots and devices. Please see the workshop webpage for details.
Please join the crescendo of ecstatic congratulations to Dr. Anurag Koul, a former postdoc in our lab and a close collaborator, who is transitioning his academic career into an applied scientist role at Amazon NYC. Bravissimo!
Raising a cheer of congratulations to Dr. Shaoru Chen, a former postdoc in our lab and collaborator, who recently sold his soul to LinkedIn joined LinkedIn!
Organizing an MSR-wide workshop in Redmond, WA for folks working on group theory applications for model and policy and representation in modern learning algorithms.
Recent presentation at Yale University based on our mixed H2/H-infinity stochastic RL policy optimization analysis.
Spoke at Microsoft Research Game Intelligence Group in Cambridge, UK on Dynamics from Pixels as Time Integrations of Neural Lie Group Homomorphisms. Slides
New robust policy optimization paper, based on H-infinity control principles for accelerating the convergence of traditional policy optimization schemes in infinite-horizon control settings out in IFAC world congress 2023.
Serving as Associate Editor for IEEE's International Conference on Robotics and Automation (ICRA) Workshops.
Delivered a talk at Microsoft Research NYC on A Short Treatise on the Kinematics and Kinetics of Robots. Slides
Invited to serve as Associate Editor for IEEE's International Conference on Robotics and Automation (ICRA).
Our new soft robot mechanism for motion correction in emerging MRI-LINAC RT systems got accepted for a presentation at the joint John R. Cameron-J.R. Cunningham Young Investigators' Symposium! Please e-attend my talk if you can at AAPM's annual meeting this year.
I am teaching a robot manipulation course at Brandeis in the winter and spring of 2020 as an Adjunct Faculty. I am keeping the lecture notes here as I develop them and the course progreses.
The docker repo for my IROS 2018 Minimax iDG Submission has moved to the following repo: lakehanne/ros. Look for the idg_iros18 tag. I will continue hosting the youbotbuntu14 repo on my hub for the next six months before I finally delete it.
I have successfully defended my PhD Thesis. Thanks to the mentors I worked with during my graduate program, viz., Steve B. Jiang, and Nick Gans, and my dissertation committee members: Mark Spong, T. Summers, Y. Tadesse and D. Bhatia.
Starting in the summer of 2019, after I shall have defended my PhD thesis, I shall be resuming as a postdoc at The University of Chicago's Pritzker School of Medicine.
Thanks to the generous support of Google AI, I will be attending the full tutorials, workshops, and program sessions at NIPS 2018. Bonjour Montréal!
I had a reyt good great time at this year's IROS and ROSCon in Madrid. Here are links to pictures of my IROS talk, ROSCon meeting, and video of Marc Raibert's mini-spot demo
Deep BOO! gets accepted to the Workshop on the Algorithm Foundations of Robotics (WAFR). WAFR has an "established reputation as one, if not the most, important venue for presenting algorithmic work related to robotics".
At the Department of Brain Robot Interface (BRI) of Japan's ATR Computational Neuroscience (CNS) Labs., I presented our minimax iterative dynamic game work.
I had the pleasure of presenting my research outlook on stably learning the dynamics of nonlinear robot trajectories to the management team of Preferred Networks, Tokyo earlier this afternoon.
Our IROS2018 submission is a definite accept. Codes here. Videos on my home page. Hallo Brisbane!
Our Minimax iDG paper gets accepted to the Machine Learning for Planning and Control Workshop at ICRA 2018. Codes here. Videos on my home page. Hallo Brisbane!
I will be pitching my start-up idea to VCs in Vancouver in October. Yeah, sure. Come in and argue!
The camera-ready version of Soft-NeuroAdapt is now on arxiv. Codes are on github.
Awarded the NSF Doctoral Consortium Award.
Awarded the Open Software for Robotics Foundation (OSRF) scholarship.
Our submission got accepted to IROS abstract only track. Video here.
Brandon Amos' OptNet paper was accepted to ICML. I am mentioned in the acknowledgement section alongside Ian Goodfellow.