Pinned Tweet
We are honored to share that Super Odometry is now published in @ScienceRobotics and featured as a highlight article! 🚀 This work rethinks the SLAM paradigm: true resilience should not rely solely on external perception—it should begin from within.
science.org/stoken/author-to… #SLAM
Feel free to join us tomorrow for the Tartan IMU Challenge Presentations!
We received submissions from 131 teams and selected the Top 10 teams to present their solutions and share their approaches.
Zoom: cmu.zoom.us/j/7802647225?omn…
Robots can see the world. But can they sense themselves?
Join us online tomorrow at #IROS2026 for:
Beyond Exteroception: Interoceptive Perception for Resilient Robotics
Speakers include Davide Scaramuzza, Maani Ghaffari, Chen Feng, Carmelo Sferrazza, Haozhi Qi, Daniel Gehrig, Yuheng Qiu, Wenshan Wang, and Shibo Zhao.
We’ll discuss IMU learning, proprioception, humanoids, state estimation, sensor fusion, and embodied intelligence.
📅 Sept. 27
⏰ 8:40 AM–5:00 PM EDT
💻 Join online: superodometry.com/interocept…
#IROS2026 #Robotics #EmbodiedAI #HumanoidRobotics #IMU #Proprioception
Happy to share that the TartanIMU Challenge has reached 54 teams with a total of 745 submissions. The competition will officially conclude on September 20th. We look forward to seeing your final solutions ^^
Please check the details on our challenge:
kaggle.com/competitions/tart…
Robots can see the world.
But can they sense themselves?
Humans don't rely solely on vision. Even in darkness, we continuously perceive motion through vestibular and proprioceptive sensing.
Today we have foundation models that understand images, language, and video. Perhaps the next frontier is a foundation model that helps robots sense themselves.
We’re honored to open the IROS 2026 Learning IMU Odometry Challenge.
Workshop: superodometry.com/interocept…
Challenge: superodometry.com/imuchallen…
TartanIMU: superodometry.com/tartanimu
Discord: discord.gg/Huf2GJx32y
#Robotics #IROS2026 #EmbodiedAI #RobotLearning #FoundationModels #Proprioception #IMU #Humanoid #SLAM
That is pretty cool! Curious about your thoughts on using LiDAR or depth image on perception.
One million hours of human video is pretty impressive.
Today we are introducing Dyna-2, a world-action model pre-trained on one million hours of human video. At this scale, for the first time, we discovered several new scaling laws:
• world-action models exhibit scaling law on human data across four orders of magnitude, from 1000 to 1,000,000 hours,
• this human data scaling law implied a scaling law on never seen robot data,
• both data and objective matter; world modeling and scaling on video data are essential for cross-embodiment scaling transfer to emerge
🧵
Robots can see the world.
But can they sense themselves?
Humans don't rely solely on vision. Even in darkness, we continuously perceive motion through vestibular and proprioceptive sensing.
Today we have foundation models that understand images, language, and video. Perhaps the next frontier is a foundation model that helps robots sense themselves.
We’re honored to open the IROS 2026 Learning IMU Odometry Challenge.
Workshop: superodometry.com/interocept…
Challenge: superodometry.com/imuchallen…
TartanIMU: superodometry.com/tartanimu
Discord: discord.gg/Huf2GJx32y
#Robotics #IROS2026 #EmbodiedAI #RobotLearning #FoundationModels #Proprioception #IMU #Humanoid #SLAM
[7/n] A sincere thank you to everyone who made this workshop, challenge, and release possible! 🙏
🔹 Corresponding Organizers: @GuanyaShi , Wenshan Wang, Shibo Zhao
🔹 Main Organizers: @MuqingC58595, @QiuYuhengQiu, @YuanJunbin , Sifan Zhou, @DrChenWang , @aero_gjy, @smash0190
🔹 Main Contributors: @YiZhaoJasper, @MaximY33479, Haomin Wen, @YutianChen03, Chris Shi, @mahaticode, @ShawnFei331
We deeply appreciate everyone’s hard work across the benchmark, dataset release, website, and outreach!
[5/n] Our long-term vision goes far beyond IMU odometry. We hope to build an open community for proprioception foundation models, bringing together researchers across robotics to share datasets, benchmarks, models, and ideas spanning IMUs, proprioception, and other body-centric modalities.
If you're interested in contributing data, models, or ideas, we'd love to have you join us at IROS 2026 and become part of this community.
Workshop: superodometry.com/interocept…
Challenge: superodometry.com/imuchallen…
TartanIMU: superodometry.com/tartanimu
Discord: discord.gg/Huf2GJx32y
[4/n] Alongside the challenge, we are excited to release the TartanIMU codebase.
It represents our initial attempt to learn shared physical representations across embodiments to predict motion velocity across diverse platforms. While zero-shot generalization to unseen systems remains an unsolved hurdle, we share this work as a humble starting point and baseline for the community.
The video below demonstrates TartanIMU on drone data. Results across other platforms and full code can be found here:
📦 Code: github.com/superxslam/Tartan…
🌐 Project: superodometry.com/tartanimu
[3/n] To accelerate research in proprioception, we’re launching the Learning IMU Odometry Challenge! All data is fully prepared and ready for training.
The Goal: Predict 3D body velocity directly from raw IMU signals across four diverse platforms: UGVs, Drones, Quadrupeds, and Handheld setups. Choose your path: a unified generalist model or tailored expert models.
Top contributions will receive the Best Paper Award at our IROS workshop! 🏆
Challenge: superodometry.com/imuchallen…
Workshop: superodometry.com/interocept…
[2/n] "Proprioception and tactile sensing are as important as vision for intelligent robots." — Prof. Jitendra Malik. @JitendraMalikCV
While foundation models have transformed how machines process images, video, and language, Proprioception remains largely unexplored. Because IMUs are the universal heartbeat of robotic self-sensing, we ask: can machines learn an innate, resilient sense of motion from within?
[1/n] Over the past decade, robotics has made remarkable progress in cameras, LiDAR, and vision foundation models. Yet these sensors often fail together in smoke, darkness, dust, occlusion, or rapid motion.
Biological systems solve this differently: they first sense themselves, then understand the world.
Excited to present our work "SuperMap: A Spatio-Temporal SLAM System for Visual-Language Navigation" at RSS 2026!
SuperMap is a living spatial memory for embodied AI — it perceives the world, remembers its evolution, and supports reasoning and action.
Perceive → Remember → Reason → Act
🌍 Project Page: superodometry.com/supermap
#RSS2026 #Robotics #SLAM #EmbodiedAI #SpatialAI #CMURobotics
[9/9]Acknowledgments
This work would not have been possible without the incredible support of our collaborators.
We are deeply grateful to CMU AirLab (Prof. Wenshan and Prof. Sebastian Scherer) and Prof. Ji Zhang for extensively validating SuperOdometry and SuperMap in their visual-language navigation system. Their thoughtful feedback and continuous support were instrumental in improving the systems.
Also, Special thanks to our amazing team:
@Lukecf1 (co-lead), @ZhuAdrien, @zl3466, Changwei Yao, Nader Zantout, @seungchankim25 for their invaluable contributions to the development, deployment, and evaluation of SuperMap.
We also thank @Jonatha11563555 for helping run the large-scale experiments. 🙌
[8/9] 🌍 Vision
We believe a map should not only tell robots where the world is. It should help robots perceive the world, remember their evolution, and support reasoning and action.
That's the vision behind SuperMap.
📄 Paper: roboticsproceedings.org/rss2…
💻 Project, Interactive Demo & Code: superodometry.com/supermap
[7/9] 📊 Highlights
✔️ Fully open source with an interactive 3D map. Please try our interactive 3D instances demo ^^.
✔️ Continuous 2-hour CMU campus deployment from indoor → outdoor without retraining
✔️ Real-time & fully onboard
✔️ Persistent object identities under occlusion and scene changes
✔️ Tracks object appearance, disappearance, and relocation
superodometry.com/supermap
[6/9] 🧩 Model-Agnostic Spatial "Memory"
Rather than replacing vision foundation models, SuperMap complements them by providing the missing spatial "memory".
It seamlessly integrates modern perception models such as SAM2, Grounding DINO, Boxer, OWLv2, and more, transforming frame-level predictions into a unified, persistent, and structured world representation for downstream embodied AI.
🤖 Made with AI
[5/9] 🏗️ System Architecture
How do we build a living spatial memory?
SuperMap is organized into three layers:
1. Geometric Layer → reconstruct the world
2. Instance Layer → maintain persistent object identities
3. Topological Layer → reason over space, semantics, and time
Together, they transform perception into memory.