Huge thanks to @spaceandtech_ for featuring E-dongkit! As a research member Spatial Intelligence and System Integration, I’m glad to see our progress being shared. Our team is dedicated to achieving Robust Field Autonomy for legged robots—ensuring they can navigate the most unpredictable, off-road environments with precision. Building robots that can truly "see" and "act" in the wild is a massive challenge, but it's one we tackle with passion every day. Stay tuned as we continue to push the boundaries of what legged robots can do!

Today's humanoid robots are beasts in controlled environments. You can see that in our posts. Send one humanoid out to buy groceries and it falls apart. It can't reliably climb your stairs, cross a cracked sidewalk, or hold a line going downhill. Most of that gap isn't the hardware. It's the control policy. These policies are reactive: the robot maps what it sees right now straight to an action. No model of what its own body is about to do next. ParkourFormer goes straight at this. The idea in plain terms: before acting, it predicts its own body state a couple of frames ahead. Then it picks the action based on that forecast. Reaction turns into anticipation. A small prediction head forecasts the next two proprioceptive states: joint angles, velocities, balance. Those predictions get fed directly into the action head. The robot moves based on where it expects to land next. Not just what's under it right now. How much does this matter? Remove the supervised future-prediction loss in training and descending-stairs success collapses from 95% to under 10%. When you can't see the ground under your next step, anticipation is the whole game. The headline numbers, on a Unitree G1 with 29 joints, across nine terrains: 93.85% average traversal success. Up to 42.73% above strong MLP, MoE, and vanilla Transformer baselines. The biggest margins show up on the hardest terrains. One unified policy for stairs, gaps, slopes, rough ground, and obstacles. No per-terrain tuning. The honest caveat: this is still staged terrain, and it leans hard on RGB-D depth. Cut the depth feed and the policy fails completely. So not grocery-run ready. But "predict your next body state, then act" is the kind of shift that actually narrows the lab-to-street gap.

🤷What if we want to learn from human data... without human data? In our work NIL (No-data Imitation Learning) @CVPR, we explore a simple but ambitious question: Can robots learn directly from AI-generated videos without any curated demonstration data? 🔗nil.is.tue.mpg.de/

Holy crap atarilab.github.io/motiondis…

AME2 has got two follow-ups which can significantly improve training efficiency, performance, and generalization. I think the training cost can be reduced to (not by) 20% now. 1) active gaze arxiv.org/abs/2606.05880 2) sparsify attention arxiv.org/abs/2606.00637

Best Conference Paper Award and the Best Paper Award on Robot Manipulation and Locomotion at #ICRA2026!

Nice new blog post by David Rosen on certifiable factor graph optimization in GTSAM: gtsam.org/2026/06/02/certifi… Together with the chordal SDP work I posted about recently, this should soon lead to some very cool new certifiable optimization features in GTSAM!!!

#ICRA2026 paper "DroneKey : A Size Prior-free Method and New Benchmark for Drone 3D Pose Estimation from Sequential Images" @bini_und, Yeong-Jun Cho abs: arxiv.org/abs/2602.06211

Performance Upgraded. Possibilities Expanded. Our DR02 humanoid robot continues to evolve with enhanced payload capacity and obstacle-crossing capabilities, unlocking greater potential for real-world industry applications !

Cool work led by @ctki49 @mxu_cg @KehanWen170077 collaboration between @leggedrobotics @xbpeng4 youtube.com/watch?v=OY61TiDo…

wholebodylocomotion.github.i… In this work we explored how to repurpose terrain-aware human locomotion skills in a scalable and versatile way: like the action-free world model, we predict how humans will move online, and then let the humanoid track it with RL.

9:00-10:30 Hall C

「特徴の欠如」と「地形条件変化」に頑強な LiDAR-IMU-Leg Odometry を明日の ICRA2026 にて発表します。 @k_koide3 の元での、ラストの修行成果です。 takuokawara.github.io/RAL202… #robot

Equivariant Filter for Radar-Inertial Odometry Giulio Delama, Jan Michalczyk, Morten Nissov, Martin Scheiber, Alessandro Fornasier, Kostas Alexis, Stephan Weiss arxiv.org/abs/2604.23033 [𝚌𝚜.𝚁𝙾]

We talk a lot about how deep learning can enable new robot capabilities but a lot less about how we can use it to make robot capabilities more resilient and therefore make robots more deployable Slide from an interesting talk by Prof Kostas Alexis in robots in the wild workshop at icra

At #ICRA2026, certifiable optimization is having a moment. Our workshop paper with Avinash Subramanian, Connor Holmes, Tim Barfoot & Frederike Dümbgen brings chordal sparsity Bayes trees to globally optimal factor-graph estimation in GTSAM. Blog post and arxiv link in the replies.

We are arriving in ICRA, Vienna for presenting our Behavior Foundation Model for Humanoid Robots at Interactive Session 1 (Hall C), TuI1I.120, June 2nd. Come and strike up a conversation on anything about humanoid robots and beyond~

ヒューマノイドに「人間のように走らせる」新手法が、人間の動きを手本にせず実現する形で発表されました。 歩行はキネマティック参照（人間の歩き方データ）を使って学習させるのが主流ですが、走行は参照データが乏しく安定維持が難しい課題でした。新手法『SPRINT』は、参照動作なしで、走る動きの周波数特性を踏まえた事前情報だけからロボットが自分で安定スプリントを学ぶ枠組みです。シミュレーション・実機の両方で人間に近い疾走パターンを確認しました。 ヒューマノイドの能力が「歩く」から「走る・運動する」へ広がる基礎研究です。災害現場や警備など、機動性が要求される実用域への適用が見えてきます。

Excited to present our work, “Tightly-Coupled Dynamic Object Tracking and RGB-D Inertial Odometry Estimation with Dual Quadrics,” at #ICRA2026！ Our method enables robust odometry in highly dynamic environments. Thanks to my advisors @k_koide3, @MR2T_AIST and Prof. Jun Miura.

Proud to share our #ICRA2026 work ScaleMaster, led by HyoSeok Ju. We ask: have we really mastered scale in deep monocular visual SLAM? 📍 Wed Jun 3, 15:00–16:30 Hall C · WeI2I.167 scalemaster-dataset.github.i…