Check out our 2nd paper in "Robot Learning Science"! Key question we aim to answer: Are robot actions unique among other AI modalities in that they must be predicted as continuous values, or can discrete action representations compete with, or even outperform, their continuous counterparts? (Perhaps) surprising fundings: 1. Even with the simplest discretion action representation (512 bins for each action dimension), discrete actions match or beat continuous actions, for both single-task policies or generalist policies, in both sim and real, in both single-step and diffusion/flow settings. 2. Continuous regression/flow policies tend to over-smooth action chunks. In RoboCasa, we found that the jerk (describe how fast acceleration changes) of action chunks predicted by continuous regression is only 7% of expert actions'. The number for discrete method is a lot higher. In other words, discrete actions keep expert's temporal variation better. 3. Precision is not the main issue of discrete methods! Even for one-step discrete regression, the end-effector prediction accuracy is similar to continuous methods', actually even better on fine motions. 4. Instead, the main issue of discrete method is factorization error, which can be fixed by iterative computing (guidance or diffusion/flow or AR). Meanwhile, we found that smarter tokenizers cannot replace iterative computing.

How far can coding agents take us toward general-purpose robots? Embodied First Exam lets us study this across 1K+ tasks—from spatial puzzles to agile whole-body control and real-time interaction. Top-1 model is 33% now. Excited to help build a shared testbed for agentic robotics!

@OpenAI is so lucky to have @TairanHe99! @TairanHe99 is the first PhD student to graduate from my group, and it only took him two years and eight months to get this degree (Sep 2023 → Apr 2026). When @TairanHe99 joined my group, I thought general-purpose humanoids were an interesting research problem, rather than a solution. Humanoid whole-body control barely worked. Hardware broke everyday. Physical AGI was not even a legitimate dream. I always joke that “nothing really works in robotics,” but looking back, the community has made incredible progress—and Tairan has been a big part of that journey. I am incredibly proud of everything @TairanHe99 has achieved: @CMU_Robotics Presidential Fellowship, @nvidia Graduate Fellowship, many open-source contributions to the community, and WhynotTV podcast—one of the most influential deep tech podcasts, to name just a few. @TairanHe99’s thesis: Scalable Sim-to-Real Learning for General-Purpose Humanoid Skills. Committee members: me, @ChangliuL, @kkitani, Marco Hutter, and @pabbeel.

Congrats to @LeCARLab's 2026 #CMUGrad!

My lab at CMU @LeCARLab is hiring a postdoc! (Vibe-made the poster via @ChatGPTapp)

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 : )

We tried to rethink modern robot learning policy design in this paper. TL;DR: Generative policies perform well NOT because of their distributional-learning formulation or their ability to capture “multi-modality.” Our proposed Minimal Iterative Policy (MIP) achieves similar performance to flow with much lower inference & training cost. Some interesting takeaways: 1. “Multi-modal behavior” is a very nuanced concept in robotics. Robot learning cares about the conditional distribution P(a|o), where a is the action (chunk) and o is the observation. The objective of generative modeling in vision/language is fundamentally different from the goal in control. In vision/language, we want diverse samples from the data distribution. In control, any action that leads to better downstream performance is sufficient. Also, robotics has very little (labeled) action data, and they lie on a low-dimensional manifold. As a result, even if the marginal P(a) is multi-modal in many tasks (e.g., PushT), the conditional P(a|o) is often not multi-modal in sparse-data regions. Empirically, we observe that the common explanation—“flow/diffusion policies win because demonstrations are multi-modal”—does not hold for most studied behavior cloning benchmarks. 2. Architecture and action representation matter more than flow vs. regression. We find that the strong performance of flow/diffusion policies is largely due to their modern architectures and action representations. UNet / Transformer designs and action chunking play a vital role. In fact, L1/L2 regression policies can match flow/diffusion when paired with the right architectures and chunking, except in a few high-precision tasks. 3. What actually helps in flow/diffusion? Stochasticity injection and supervised iterative computation. To isolate these factors, we performed a “surgery” on flow to design Minimal Iterative Policy (MIP), a deterministic two-step policy that keeps only these mechanisms. Surprisingly, with far lower compute, MIP is comparable to—or even better than—flow in most studied behavior cloning benchmarks. Why? We hypothesize that these mechanisms provide good inductive biases that improve manifold adherence, especially for OOD states. Flow and MIP exhibit significantly better manifold adherence than regression. I don’t think these conclusions are complete—there are still many mysteries. But I strongly feel we need a new subfield: “the physics/science of robot learning,” aimed at understanding fundamental learning mechanisms and principles for data-driven robotics. As @ilyasut said for LLMs: “We’re moving from the age of scaling to the age of research.” For robot learning, I believe we must embrace the age of research to unlock better scaling to ignite the age of scaling.

Introduce SPIDER - a unified scalable dynamics-level retargeting method for both dex hand and humanoid! What does "dynamics-level retargeting" mean? Human motions in, physically feasible high-quality robot motions out (BOTH state and action). These robot motions are so good that you can directly open-loop execute the action sequence in the real world! Also, dynamics-level retargeting makes the downstream RL policy learning super easy since it only needs to figure out a "small residual" on top of the already physically feasible refs. Even better, we can easily do physics-level data augmentation & robustification, such as adding external forces and changing frictions. SPIDER can be used as a data engine for real2sim2real and learning from human data. How it works? It is based on the diffusion-inspired sampling-based optimal control framework we developed in the past (MBD, DIAL-MPC) and virtual contact guidance. Led by @ChaoyiPan and collaboration with @Meta FAIR.

Excited to release BFM-Zero, an unsupervised RL approach to learn humanoid Behavior Foundation Model. Existing humanoid general whole-body controllers rely on explicit motion tracking rewards, on-policy PG methods like PPO, and distillation to one policy. In contrast, BFM-Zero directly learns an effective shared latent representation that embeds motions, goals, and rewards into a common space, which enables a single policy zero-shot perform multiple tasks: (1) natural transition from any pose to any goal pose, (2) real-time motion following, (3) optimize any user-specified reward function at test time, etc. How it works? We don't give the model any specific reward in training. It builds upon recent advances in Forward-Backward (FB) models, where a latent-conditioned policy, a deep "forward dynamics model" and a deep "inverse dynamics model"are jointly learned. In such a way, the learned representation space understands humanoid dynamics and unifies different tasks. More details: lecar-lab.github.io/BFM-Zero…

100% human data -> long-horizon💡insertion on an aerial manipulator Key: Embodiment-aware diffusion policy (EADP) steers UMI's embodiment-agnostic DP using the gradient of the low-level controller's tracking error. My favorite: We quantify how "UMI-able" different robots are.

Didn't get chance to attend #CoRL2025 in person, but students from @LeCARLab will present three papers: - Sampling-based system ID for legged sim2real lecar-lab.github.io/spi-acti… at Oral5 by @NikhilSoban353 - HoldMyBeer: learning humanoid end-effector stabilization control lecar-lab.github.io/SoFTA/ at Spotlight3 by @li_yitang - Humanoid Policy ~ Human Policy human-as-robot.github.io/ at Spotlight1 by @chaitanya_cha Check them out!

On my way ✈️ to ATL for @ieee_ras_icra! @LeCARLab will present 8 conference papers (including DIAL-MPC as the Best Paper Finalist) and one RA-L paper. Details: lecar-lab.github.io/ Hope to meet old & new friends and chat about building generalist 🤖 with agility 🚀