Tianyu Li

#WorldEngine is one of the most exciting projects in AD in the past years! It's a post-training framework tackling the scarcity of long-tail safety-critical scenarios by: mining -> 3DGS reconstruction and dynamic agents control w/ behavior world models -> RL post-training.

Introducing #WorldEngine, github.com/OpenDriveLab/W…, a two-year long project. The missing infrastructure for Physical AI post-training in Autonomous driving. Open-source. Production-validated.

SimScale is selected as CVPR 2026 Oral 🎉🥳 Code, simulation data, and checkpoints have been fully open-sourced. 📷Home: opendrivelab.com/SimScale/ 📷Paper: arxiv.org/abs/2511.23369 📷Github: github.com/OpenDriveLab/S… If you find something useful, please give it a star⭐

🚀 Introducing SMASH — the world’s first high-dynamic humanoid robot for outdoor table tennis, fully autonomous with onboard perception. No motion capture needed—SMASH uses only onboard sensors for stable, full-body human–robot interaction in real-world settings. Embodied intelligence is stepping out of the lab. 🔗mmlab.hk/Smash/ #HumanoidRobots #TableTennis #RobotLearning #SportsTech #AI #OpenDriveLab #Algorithms

Humanoid robots have been prisoners of the lab. We set them free — with human data. We present EgoHumanoid: The first endorsement of human-to-humanoid transfer for whole-body loco-manipulation. 🔗 Home: opendrivelab.com/EgoHumanoid 📑 Arxiv: arxiv.org/abs/2602.10106 🧵👇

🧐Applying world models to improve real-world policy on challenging manipulation tasks used to be considered out of reach. 😌After sustained effort, we’re now seeing encouraging progress. 🚀Thrilled to introduce RISE: Self-Improving Robot Policy with Compositional World Model opendrivelab.com/kai0-rl/ arxiv.org/abs/2602.11075 RISE is, to our knowledge, the first work to use a world model as an effective learning environment for challenging real-world manipulation, enabling policy improvement on tasks that demand high dynamics, dexterity, and precision. Incredible teamwork with @lin_kunyang111 @francislee2020 @YueXiangyu @HaoZhao_AIRSUN @smch_1127

Glad to share me and my folks work of the last research cycle, good luck and have fun. Upvote on HF is appreciated! χ0: huggingface.co/papers/2602.09… χ0, a resource-efficient robotic manipulation framework addresses distributional shifts through model arithmetic, stage-aware advantage estimation, and train-deploy alignment to achieve long-horizon task reliability. RISE: huggingface.co/papers/2602.11… RISE, the first work to use a world model as an effective learning environment for challenging real-world manipulation, enabling policy improvement on tasks that demand high dynamics, dexterity, and precision. EgoHumanoid: huggingface.co/papers/2602.10… EgoHumanoid, the first endorsement of human-to-humanoid transfer for whole-body loco-manipulation. SparseVideoNav: huggingface.co/papers/2602.05… Vision-language navigation systems traditionally require detailed instructions but can be improved by incorporating video generation models with sparse future planning for faster, more efficient real-world deployment.

1/n 🎉 SimScale: Learning to Drive via Real-World Simulation at Scale 🤖 An innovative real-world simulation pipeline and a real-sim co-training strategy that significantly boost the robustness and generalization of any end-to-end planner. 📈 For the first time, we reveal the scaling effect of simulation data in autonomous driving: with zero extra real-world data, simply scaling up simulation alone can keep improving model performance!

🚀 #MTGS is now open-sourced! It manages to leverage multi-traversal data for scene reconstruction with better geometry. We utilize the nuPlan dataset with extensive multi-traversal data. 📷 github.com/OpenDriveLab/M…

🚀 #RoboDual is now OpenSource! We’re thrilled to release the code for RoboDual, one of the earliest dual-system framework for generalizable robotic manipulation that bridges the gap between VLMs’ reasoning and real-time control. 📌 GitHub: github.com/OpenDriveLab/R… 🔗 Project Page: opendrivelab.com/RoboDual/ #Robotics #EmbodiedAI #opensource 🧵[1/4]

Co-organized with @vveicao, @MHallgarten0797, Daniel Dauner, and Kashyap Chitta, with data from @motionaldrive. 🧵(3/3)

Your score depends on: 1️⃣ Performance on real frames 2️⃣ Performance on "nearest" imaginary futures to your model's past outputs, making this the *first* challenge testing robustness against causal confusion and compounding errors on real driving data 🚗 🧵(2/3)

Announcing the 2025 NAVSIM Challenge! What's new? We're testing not only on real recordings—but also imaginary futures generated from the real ones! Two rounds: #CVPR2025 and #ICCV2025. Submit in May for Round 1! opendrivelab.com/challenge2025 🧵(1/3)

@TobiasFischer11 @zhenjun_zhao @MattNiessner @francislee2020 @Haithem_Turki Aha! It seems we have missed a line of works! SUDS and ML-NSG made wonderful works with nerual raidiance field. 4DGF also has the ability to handle such data with lightning speed⚡️. We would like to disscuss them in the revised papers.

@sephy_li @zhenjun_zhao @MattNiessner @francislee2020 Thanks for your response. Consider multiple works deal with multi-traversal: SUDS, @Haithem_Turki et al., CVPR'23 ML-NSG, me et al., CVPR'24 - introduces a *very* similar scene graph idea to your multi-traversal one 4DGF, me et al., NeurIPS'24 Concurrent to ours: OpenMARS Dataset

MTGS: Multi-Traversal Gaussian Splatting @sephy_li, Yihang Qiu, Zhenhua Wu, Carl Lindström, Peng Su, @MattNiessner, @francislee2020 tl;dr:traversal->sub-graph->shared static node for background+appearance node for variations+transient node for dynamics arxiv.org/abs/2503.12552

@TobiasFischer11 @zhenjun_zhao @MattNiessner @francislee2020 Oops! I am the author of this work, and I just noticed that you and your work "Dynamic 3D Gaussian Fields for Urban Areas" are also working on the reconstruction of multiple sequence/traversal data. We will remove the related claim and also discuss your work in the paper.

@zhenjun_zhao @sephy_li @MattNiessner @francislee2020 "we propose Multi-Traversal Gaussian Splatting (MTGS), the first approach able to reconstruct multi-traversal dynamic scenes" - I beg to differ 😶

🚀 The Autonomous Grand Challenge 2025 is about to begin! This year, we’re bringing you three thrilling tracks in Embodied AI and Autonomous Driving, with a total prize pool of $100,000! Now get ready and join the competition! 🌏 World Model Challenge by @1x_tech 🛣 NAVSIM v2 End-to-End Driving Challenge 🤖 AgiBot World Challenge Visit the challenge website: opendrivelab.com/challenge2025 And more on @CVPR #CVPR2025 : opendrivelab.com/cvpr2025

Can robots learn without training❓ [𝗜𝘁'𝘀 𝗼𝗽𝗲𝗻 𝘀𝗼𝘂𝗿𝗰𝗲𝗱 ⬇ ] Teaching robots to do complex tasks WITHOUT spending hours training them. Sounds cool, right? That's exactly what DIAL-MPC does! The first training-free method for whole-body torque control using full-order dynamics: ✅ Instantly checks if a robot's moves are right or wrong ✅ Adapts quickly to new tasks without needing extra training ✅ Could work hand-in-hand with other robot learning methods Robots are getting smarter AND faster without the need for long training sessions. Website: lecar-lab.github.io/dial-mpc/ Paper: drive.google.com/file/d/1Z39MCv… Code: github.com/LeCAR-Lab/dial… Saw this first @HaoruXue

RT @OpenDriveLab: 🎉 Delighted to share that our research on driving world models, Vista, has been accepted at #NeurIPS2024 @NeurIPSConf! Ku…