Harsh Gupta @HarshGupta

With @JayHe748646 @KehanWen170077 @leggedrobotics , we push humanoid mobility to next frontier. Achieved by a generalist contact planner + a generalist tracker + a generalized neural mapping, warmstarted from few human data. Website: leggedrobotics.github.io/ACP/ Promo Video 👇

If video models scale like LLMs, they'll soon generate a perfect video plan for any embodied task. Do we need anything else? Introducing Mirror, which learns to follow these plans just by practicing, and even crafts an iron pickaxe in Minecraft w/o rewards or action labels! (1/n)

Can humanoids adapt to new terrain using only flat-ground references? 🤖 Meet P-BFM! 🥳 We learn a terrain-agnostic intent space with unsupervised RL and forward-backward representations, paired with a terrain-aware policy that adapts across diverse terrains—without terrain-specific demonstrations. 📄🎬💻 Check out our paper, video, and code here! 🌐 roboparty.github.io/P-BFM/

Most robot policies treat time implicitly, using sequence position as a proxy for physical time. But real robot systems rarely run on a fixed clock. Sensors run at different rates, observations arrive with delays, and actions take time to compute and execute. Successful control depends not only on what the robot observes but also on the timing of observations and actions. We introduce 🕒 Chronos Policy, a diffusion policy that explicitly conditions on observation capture times and requested action execution times. We use time-aware positional encodings to make attention depend on elapsed physical time. Because timestamps let the policy interpret each sequence's timing, we can randomize sensing and action timing during training without making those variations ambiguous. This exposes the policy to a wider range of timing conditions, improving robustness to timing changes at deployment.

Excited to share our work PredActor, a predictive action diffusion policy for steerable humanoid control. This direct action-generation view is appealing because it avoids the generator-tracker mismatch. 🤩

Lei Ye (Raymon Yip) @MasterYip3

2 weeks ago

Can humanoid motion generation and control be unified in a single policy? PredActor uses a single end-to-end diffusion policy rather than separate motion generation and tracking. With proprioceptive feedback, it predicts internal future states and directly outputs actions at 50

13 17 149 168K 114

🌐 Website: eigendexplore.github.io 📄 Paper: arxiv.org/abs/2610.07681 💻 Code: github.com/hgupt3/eigende… Co-led with @tylerlum23, with amazing collaborators @ChanghaoWang78 and @chuer_pan, and advisors Karen Liu, @leto__jean and @SongShuran at @StanfordAILab

Try the interactive project page: eigendexplore.github.io Visualize robot rollouts in 3D, compare policies, and explore principal components learned from human hand motion

Beyond RL, the same idea helps sampling-based trajectory optimization. With SPIDER, we retarget human hand-and-object motion to robot hands. EigenDEXplore lowers average retargeting cost by 16% across six hands, with the same sampling budget.

Per-joint Gaussian noise is a terrible way to explore a 20+ DoF hand We introduce EigenDEXplore, which explores along coordinated patterns from human hand motion, while keeping full control of every joint Better learning, from tool use 🔨 to reorientation 🔄 to bimanual 👐

EigenDEXplore also helps functional retargeting: reproducing human-demonstrated object motion with two robot hands With DexMachina's reference-guided RL, it improves average object-tracking accuracy by 28% with task reward alone

These gains transfer to the real robot On DeXtreme cube reorientation, mean consecutive successes rise from 8.45 to 11.90 (+41%) over 40 trials per method Human-guided exploration during training, full joint-space control at deployment

Cube reorientation can require motions outside a human-derived subspace Eigen-Space restricts actions to that subspace and struggles EigenDEXplore uses it to guide exploration, keeping full joint-space actions: 40% more consecutive successes than Joint-Space in sim

What changes when exploration is coordinated? EigenDEXplore discovers useful grasps and learns faster in this tool-manipulation comparison. On real-world tool use, average task progress rises from 33% to 82% across hammering, wiping, scooping, and brushing.

Dexterous manipulation requires coordinated joint motion, but standard RL often explores with independent joint noise EigenDEXplore combines independent joint noise with noise along "eigen" directions: coordination patterns extracted from human motion using PCA

How can we train better behavior from imperfect robot demos? In NEEDLEWORK, we study how we can get stitching to work on robot data, and more importantly, why these connections help. Website: needle-work.github.io Paper: arxiv.org/abs/2610.02339

Transformer Transformer will be presented at #CoRL2026 🚀 See y'all in Austin! Website: transformer-transformer.github.io Video: youtube.com/watch?v=TTyjvP… Paper: transformer-transformer.github.io/static/paper.p…

New paper release: Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids Paper arxiv.org/pdf/2608.29769 Videos nemantor.github.io/sparse-3d-trav… We study how to do learning and sim2real for 3d traversal, such as monkey bars and overhanging obstacles.

When a robot learns a new behavior, what keeps it from forgetting the ones it already knows? Prior work, especially on VLAs, shows that rehearsing past experiences can enable strong continual learning. But why does rehearsal work so well, and when does it fail? We find that a small but ✨special subset✨ of rehearsed data largely determines whether old behaviors are retained or forgotten. We call these examples Memory Anchors. ⚓ Withholding just the top 10% of Memory Anchors from rehearsal increases forgetting by up to 4.5x. Increasing their presence, meanwhile, reduces task forgetting and enables a real robot to learn challenging task sequences. Website: robot-adaptation.github.io/MemoryAnchors/ Paper: arxiv.org/abs/2608.26545 Curious? Read on! 🧵👇 (1/9)

Teleop systems are usually designed for a single embodiment, but do they have to be? Introducing ModPack 🎒: a modular teleoperation interface for bimanual mobile robots. A wearable backpack provides shared infrastructure, robot-specific leader arms adapt to different embodiments, and plug-and-play modules add capabilities like haptic feedback, active perception, and mobile manipulation. 🧵(1/n)

Robot learning is moving beyond policies built for one robot, one scene, one task. At MIT, we’re exploring a different path: turning video world models into embodiment-agnostic robot policies. Introducing VERA: a 14B video-to-action system that controls robots across embodiments, skills, and environments. From zero-shot pick-and-place on a real Panda arm to contact-rich cube reorientation with a 16-DoF robotic hand. Different robots. Different environments. Different tasks. Same video planner. Same weights. We’re open-sourcing everything so you can fine-tune VERA for your own robot setup too. Deep dive in the thread: 🔗 vera.csail.mit.edu 🧵 (1/7)