Associate Professor at Carnegie Mellon University | he/him
Pittsburgh, PA
Joined September 2011
- Tweets525
- Following627
- Followers6.8K
- Likes8.6K
David Held retweeted
Life update: I recently graduated from CMU RI, and last year I went through the academic faculty job market in robot learning. I was fortunate to receive a lot of help and support from many people throughout the process.
Compared to industry job applications, there are relatively few resources available online about the faculty application process. So I decided to open-source all of my application materials from last year here:
docs.google.com/document/d/1…
I hope these materials can be helpful to others who are considering pursuing an academic career!
David Held retweeted
Robots can’t always see the physics they act under: hidden mass, friction, kinematic constraints. A scalar reward only says "failed." The outcome of an action says much more.
Introducing SCOUT: test-time policy adaptation from action-outcome feedback. Sim-trained, adapts in the real world in a few tries 🧵👇 #CORL2026
David Held retweeted
A man in Australia asked his agent (Claude running on OpenClaw) to book him a spot in a popular gym class. The agent found a software vulnerability that let it book the class weeks further ahead than should have been possible. When the user then asked if it could move him up the waitlist, the agent discovered the API had no authorisation checks on cancelling other people’s reservations, so it cancelled the person in the first spot and moved him up the list.
Some people will call this misalignment, but his agent was perfectly aligned to him - it was only trying to help its user get what he wanted. The most important thing about this story, in my opinion, is that it gives you a window into what is about to start happening on a massive scale once millions of people have an agent trying to get their beloved users the best seats, bookings, appointments or reservations through absolutely any means necessary.
David Held retweeted
Boo! 👻 GHOST has been accepted to #RSS2026!
Can we learn manipulation skills from human video (without action retargeting)?
Yes! GHOST learns generalizable manipulation skills by training a hierarchical policy with an embodiment-agnostic sub-goal predictor + embodiment-specific controller.
🧵(1/7)
David Held retweeted
🧩 #ICML2026 💥 How can a model discover the 3D objects in a scene—their shape, color, and position—without any labels? Introducing 3D-DLP, a self-supervised object-centric model that decomposes colored 3D scenes (RGB-D and voxels) into a set of 3D latent particles.
David Held retweeted
Excited to present the first-ever 3D, generalizable, handheld tactile imaging system! 👇
Our system learns to visualize the inner structure of soft objects purely through touch -
"More with LESS: Local Scene Representation for Tactile Imaging," has been accepted to #RSS2026 🇦🇺
Curious how we leverage the locality of touch for large-scale tactile SSL? Read on! 🧵
David Held retweeted
My first last-author paper, with @lakshitadodeja and @rai_inst collaborators!
When deploying Behavior Cloning (BC) policies, we encounter performance gaps due to compounding error, subtle task shifts, etc.
We want to be able to post-train BC policies with a few online interactions to quickly improve performance.
Q2RL is what we hoped for in a robot post-training algorithm:
- ✅it starts with BC policy rollouts (no need for large, diverse datasets), and
- ✅it's sample-efficient enough for real-world online reinforcement learning!
See the thread and linked website for more details. Paper and code are released!
Can BC policies be quickly improved through real world experience?
Our new #RSS2026 paper proposes Q2RL, a method that bridges BC and RL for on-robot learning.
Q2RL improves BC policies by up to 3.75x with just 1-2 hours of online interaction!
So when life gives you BC, make Q-functions! 🍋
Details in thread 🧵
David Held retweeted
I’m attending #ICLR2026 in Rio this week to present LPWM! Friday April 24 Poster Session 3 10:30, Oral Session 4B 3:15p. Happy to chat about self-sup object-centric learning and world models. I’ll be on the job market soon and looking for exciting opportunities!
#ICLR @iclr_conf
🚀 #ICLR2026 Oral 💥
How can we design world models that capture object interactions directly from pixels?
Introducing Latent Particle World Models-the first end-to-end self-supervised, object-centric world model, trained from videos, supporting action/img/lang conditioning.
1/n
David Held retweeted
I’m so tired of writing rebuttals to this kind of “lack of novelty” review: “This paper trivially combines A, B, and C, so the algorithmic novelty is limited.”
Technically, most (if not all) robotics papers are convex combinations of existing ideas.
I still deeply appreciate A+B+C papers—especially when they deliver:
- New capabilities: the “trivial combination” unlocks behaviors we simply couldn’t achieve before
- Sensible & organic design: A+B+C is clearly the right composition—not some arbitrary A′+B+C′
- Nontrivial interactions: careful analysis of the dynamics, coupling, or failure modes between A, B, C
- Rehabilitating old ideas: A was dismissed for years, but paired with modern B/C, it suddenly works—and teaches us why
- System-level & "interface" insight: the contribution is not any single piece, but how the pieces talk to each other
- Scaling laws or regimes: identifying when/why A+B+C works (and when it doesn’t)
- Engineering clarity: making something actually work robustly in the real world is not “trivial”
- New problem formulations: sometimes the real novelty is in the reformulation—only under this view does A+B+C make sense.
Maybe worth keeping these in mind when reviewing the next A+B+C paper : )
Excellent analysis by @Majumdar_Ani, recommended read!
David Held retweeted
🚀 #ICLR2026 Oral 💥
How can we design world models that capture object interactions directly from pixels?
Introducing Latent Particle World Models-the first end-to-end self-supervised, object-centric world model, trained from videos, supporting action/img/lang conditioning.
1/n
I am filling out a fellowship application letter of recommendation, and one question is: "Applicant's achievements reflect his or her ability: * - select -Not at All; Well; Moderately Well; Very Well; Extremely Well".
If I select "Well" then what does that imply?
What are the best slides you have seen to explain transformers to someone who has never seen them before?
David Held retweeted
Introducing FMVP: a method that adapts to natural arm motions during robot-assisted dressing.
Pre-trained on vision in sim, fine-tuned with limited real-world vision+force data, and tested in a 12-user, 264-trial study, FMVP is robust across garments and motions. #CoRL2025
Check out our new CoRL paper on how to find shapes that cause failures for robot manipulation policies!
How do we discover a robot's failure modes before deploying it in the real world? Standard benchmarks often don't capture the full picture, leaving policies vulnerable to plausible variations in object shape.
Thrilled that our work, "Geometric Red-Teaming for Robotic Manipulation," has been accepted as an oral presentation at #CoRL2025! We introduce a framework to automatically find these geometric blindspots.
georedteam.github.io
🧵
Check out our CoRL 2025 paper on learning from past interactions; we show, theoretically and experimentally, the benefit of using a generator-verifier paradigm!
A closed door looks the same whether it pushes or pulls. Two identical-looking boxes might have different center of mass. How should robots act when a single visual observation isn't enough?
Introducing HAVE 🤖, our method that reasons about past interactions online! #CORL2025
Check out our new CoRL paper on planning in continuous action spaces by searching over sequences of object transformations!
🚨Introducing SPOT: Search over Point Cloud Object Transformations. SPOT is a combined learning-and-planning approach that searches in the space of object transformations.
Website: planning-from-point-clouds.g…
Paper: arxiv.org/abs/2509.04645
Code: github.com/kallol-saha/SPOT
Check out our latest work on opening unseen articulated objects in the real world; this is a single policy that operates zero-shot on 3 robot embodiments!
Key insights:
1) Large-scale sim training
2) Hierarchical imitation learning
Introducing ArticuBot🤖at #RSS2025, in which we learn a single policy for manipulating diverse articulated objects across 3 robot embodiments in different labs, kitchens & lounges, achieved via large-scale simulation and hierarchical imitation learning.
articubot.github.io
🧵
Does anyone have a copy of the ShapeNetSem dataset? The Sshapenet website is having issues for downloading the dataset. Thanks!!
A colleague mentioned this to me today and I just checked it: ChatGPT still cannot count the number of r's in strrrawberrrrry:
chatgpt.com/share/67dd8bef-2…
