Hydrobotics

@shi_weiyan @DL4Code Congrats!

Excited to share that our paper won Best Paper 🏆 at the @DL4Code workshop at ICML: "Coding with Enemy: Can Human Developers Detect AI Agent Sabotage?" This was a challenging project (100+ developers, 5-hour+ coding sessions, 10-month effort) but it sends an important message: AI safety isn't just about aligning models, it's also a human-AI problem. tldr: (1) When a coding agent has a side task (e.g., inserting malicious code), 94% of developers fail to detect it. (2) Even when a monitor flagged the malicious code, 63% (12/19) approved it anyway, because they didn't fully understand the large codebase and overtrusted the agent. (3) So monitor design has to account for human factors. Participants preferred proactive intervention (e.g., a concrete fix, detailed analysis, etc) over flag-only alerts. Let's make AI safety more human-centric! 💪💪💪

What makes egocentric video data critical for robot learning? Frontier robotics and world-model research demands massive volumes of first-person video showing natural human behavior in diverse real-world environments. Egocentric video data captures the world from the perspective of the actor performing a task, providing the visual grounding that embodied AI models need to connect perception with action. Unlike third-person footage, first-person video preserves the spatial relationships between hands, tools, and objects as they appear during manipulation. #PhysicalAI

@1x_tech Alien tek indeed

NEO’s Hands An API to the Physical World

Introducing the Global Researcher Map 🌎 We mapped every AI researcher into a visual landscape you can explore Search your favorite authors, topics, or institutions, and see who’s behind the work

Simply mixing raw, in-the-wild human videos into robot training actually hurts policy performance. The fundamental problem is that humans and robots move in fundamentally different ways. EgoWAM closes this embodiment gap by focusing on world representations instead of raw motions👍🏻 Motion-centric representations and strong visual pretraining each tackle different parts of this gap.

@BaoyuLi6 Nice work! Naive BC co-training on raw human motions backfiring due to embodiment mismatch is a super important gotcha. Looking forward to the code & data drop.

Egocentric human data is abundant, but human motion is not always positive supervision for robot policy due to embodiment gaps. Naive BC co-training can HURT performance ☹️. 🌟Our key finding in **EgoWAM**: the state-prediction branch of a World Action Model effectively bridges this embodiment gap, enabling robot performance to scale with diverse **in-the-wild** human data. 💡The key question then becomes: what world representation transfers best across embodiments? 👇🏻Let’s take a deep dive into it: 🌐 gatech-rl2.github.io/egowam.github.… 🧵[1/]

@aymenmmir Congrats!

1/4 📢 AHOY! accepted at #ECCV2026! 🎉 From a single in-the-wild YouTube video, often heavily occluded, AHOY builds a complete, animatable 3D human avatar. Each clip below runs input → 360° avatar → animation. 🧵👇 🔗 miraymen.github.io/ahoy/

Every video passes through an automated quality pipeline that detects issues and removes non-compliant frames before delivery. - Hand detection & visibility scoring - Auto-removal of frames without hand presence - Corrupt frame and dropout detection - Low-light & excessive motion filtering - Audio quality flags - Privacy compliance verification

The primary bottleneck for robots in real-world settings, particularly homes, is not hardware, but high-quality, diverse human action data. Traditionally, training robots required engineers to manually teleoperate arms in labs. This approach was costly, slow, and produced limited, poorly generalizable data. Hydrobotics offers a superior solution. We use iPhone motion sensors to crowdsource authentic videos of everyday human tasks such as washing dishes, walking, and folding clothes. Our AI-powered quality validation pipeline then converts these into structured action sequences containing speed, trajectory, acceleration, and force data—rich "motion textbooks" that robots can directly learn from. By reducing collection costs by dozens of times while delivering 50-70% gross margins, we provide scalable, high-fidelity data far beyond traditional methods. Hydrobotics is building the essential action data infrastructure to accelerate embodied AI from lab prototypes to practical deployment. #Hydrobotics #PhysicalAI

Physical AI just got its summit. And this is only the beginning.

What if robots could remember and learn from their own fastest success, even when it came from a “lucky” trial? People often treat efficiency as something to optimize after success. Our new work, Temporal Self-Imitation Learning (TSIL), takes a different view: fastest success itself can be a useful training signal in reinforcement learning. TSIL turns rare fast successes discovered during interaction into two learning signals: adaptive temporal targets that encourage faster completion in a self-paced way, and fast-success self-imitation that helps preserve efficient behaviors before they are forgotten. Across 15 challenging long-horizon manipulation tasks, TSIL improves learning efficiency, task-completion efficiency, and training stability. If training your robot policy feels like endless tuning, give TSIL a shot😄! Grateful to my advisor @Boyuan__Chen for the guidance and support on this work. - Paper: arxiv.org/abs/2606.19752 - Project page: generalroboticslab.com/TSIL - GitHub: github.com/generalrobotic…

@virtuals_io @eastworlds_io @UnitreeRobotics A meetup by the sea 🌊 #RSS2026

The people who build robots and the people who train them are coming together in one room. Join us for an evening co-hosted by @eastworlds_io and @UnitreeRobotics at RSS 2026 in Sydney. RSVP below.

What happens when robot world models learn from human experience at scale? 🤔 DreamDojo from NVIDIA Research is a generalist robot world model pretrained on 44K hours of egocentric human videos and then post-trained on robot data to generalize across new objects and environments. After distillation, it runs at 10 FPS for live teleoperation, policy evaluation, and model-based planning. Read the ICML paper to learn more 📄 nvda.ws/3TlTCaw

@ShenyuanGao Would love to connect!

Our poster is at📍Hall A #3815, 5:00-6:45 PM KST, Wednesday, July 8. I will be on site from July 6-11 and welcome connect!

.@NVIDIARobotics has just released a fine-tuned GR00T N1.7 robot policy on Hugging Face. It was trained on the LIBERO benchmark for Panda robot arm manipulation tasks.

@HoucineOusmer @MACHINASUMMIT @joinstationf See you!

🚀 Very exciting to participate at @machinasummit 2026 at @joinstationf Paris! let’s connect, debate, and map out the future together at Station F. See you there! 🤝 #MachinaSummit #Physicalai

@GaotangLi RubricEM looks like a clean instantiation of the first one turning rubrics into structured, stage-aware feedback for longhorizon agents

“Three ways to the unverifiable: Rubrics as Rewards, Generative Reward Models, Process Rewards” Check out how we instantiate the exact same idea in arxiv.org/abs/2605.10899!

@AdhamGhazali Make sense

Lots of people are entering robotics by skipping the foundations. The pitch is that world models understand physics so you don't need them. Robot physics was solved decades ago. Calibration, motion planning, collision avoidance. Known quantities, not open research. I get the appeal. Train one model, let it handle calibration, motion, collisions, and the task itself. But watch the demos these teams ship. Autonomous footage played back at 4-10x. That's not a hardware limit. That's a team without foundations. Strong foundations allow you to show your videos at 1x proudly.

Robotics is hitting a wall that larger models alone cannot climb. Physical AI cannot scale on videos alone, and these papers show why. @ManycoreTech research team’s #ECCV2026 accepted paper SPEAR turns Unreal Engine from a visual simulator into a programmable robotics training system. Its big step is exposing 14K UE functions to Python while rendering 1080p frames at 73fps. Gives you scriptable worlds, agents, cameras, materials, labels, and deterministic scene execution. #Simulation #EmbodiedAI #SpatialIntelligence