𝗔𝗳𝘁𝗲𝗿 𝟭𝟬+ 𝘆𝗲𝗮𝗿𝘀 𝗶𝗻 𝗿𝗼𝗯𝗼𝘁 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴, from my PhD at Imperial to Berkeley to building the Dyson Robot Learning Lab, one frustration kept hitting me: 𝗪𝗵𝘆 𝗱𝗼 𝗜 𝗵𝗮𝘃𝗲 𝘁𝗼 𝗿𝗲𝗯𝘂𝗶𝗹𝗱 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗼𝘃𝗲𝗿 𝗮𝗻𝗱 𝗼𝘃𝗲𝗿 𝗮𝗴𝗮𝗶𝗻? 𝗧𝗵𝗲 𝗽𝗮𝘁𝘁𝗲𝗿𝗻 𝗜 𝗸𝗲𝗽𝘁 𝘀𝗲𝗲𝗶𝗻𝗴: • New robotics team starts • Spends 6 months building data collection pipeline • Spends another 3 months debugging synchronization issues • Finally starts collecting task-specific data • Realizes their infrastructure choices limit their flexibility • Starts over 𝗧𝗵𝗶𝘀 𝗶𝘀 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝗽𝗼𝗶𝗻𝘁 𝗼𝗳 𝗿𝗼𝗯𝗼𝘁 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴: Robot learning is fundamentally data-driven. Whether you're picking strawberries or assembling electronics, the core infrastructure needs are identical. That's actually why I was so interested in pursuing data-driven robotics over a decade ago. 𝗬𝗼𝘂 𝗮𝗹𝘄𝗮𝘆𝘀 𝗻𝗲𝗲𝗱: • Multi-sensor data synchronization across different frequencies • Flexible storage that works with future algorithms • Visualization tools to understand your data • The ability to experiment with different temporal resolutions • Robust logging that captures everything you might need later The trend towards AI in robotics is growing, with robots needing to process and analyze large amounts of sensor data to manage variability and unpredictability in real environments. 𝗕𝘂𝘁 𝗲𝘃𝗲𝗿𝘆 𝘁𝗲𝗮𝗺 𝗯𝘂𝗶𝗹𝗱𝘀 𝘁𝗵𝗶𝘀 𝗳𝗿𝗼𝗺 𝘀𝗰𝗿𝗮𝘁𝗰𝗵. Imagine if every web developer had to build their own database, web server, and deployment pipeline before writing their first line of application code. 𝗧𝗵𝗶𝘀 𝗶𝘀 𝘄𝗵𝘆 𝗜 𝗳𝗼𝘂𝗻𝗱𝗲𝗱 𝗡𝗲𝘂𝗿𝗮𝗰𝗼𝗿𝗲. Instead of every robotics team spending months on infrastructure, we provide the common tools that let you go from "I have a robot" to "I'm shipping intelligent robot behaviors" in days, not months. 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗶𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻 𝗶𝗻 𝗿𝗼𝗯𝗼𝘁𝗶𝗰𝘀 𝘄𝗼𝗻'𝘁 𝗰𝗼𝗺𝗲 𝗳𝗿𝗼𝗺 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝗿𝗲𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗽𝗹𝘂𝗺𝗯𝗶𝗻𝗴. 𝗜𝘁'𝗹𝗹 𝗰𝗼𝗺𝗲 𝗳𝗿𝗼𝗺 𝘁𝗲𝗮𝗺𝘀 𝘄𝗵𝗼 𝗰𝗮𝗻 𝗳𝗼𝗰𝘂𝘀 𝗲𝗻𝘁𝗶𝗿𝗲𝗹𝘆 𝗼𝗻 𝘄𝗵𝗮𝘁 𝗺𝗮𝗸𝗲𝘀 𝘁𝗵𝗲𝗶𝗿 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝘂𝗻𝗶𝗾𝘂𝗲. Robot learning shouldn't be bottlenecked by infrastructure. It should be bottlenecked by creativity. What's the longest you've spent building infrastructure before getting to the actual robotics problem you wanted to solve?

Two big things here: 1) BiGym 2.0, a new whole-body humanoid benchmark. 2) The first thorough comparison of learned policies vs LLM-written robot policies. I was surprised how far coding agents got from just 1 demo, where learned policies would be useless. But ultimately, learned policies (from scratch or VLAs) are the more general solution, handling a broader range of tasks, albeit with more demos.

Wrist cameras help robots a lot. Taking them off and letting the robot look at the right thing helps more. EyeRobot 2.0, from Kush Hari, @justkerrding and the team at @UCBerkeley and Amazon FAR (advised by @Ken_Goldberg and @akanazawa), has no wrist cameras at all. One stereo head moves its gaze to whatever matters at each step. Sharp in the middle, blurry at the edges, like our own eyes. Seven real tasks, including capping a marker, zipping a bag and pulling a pan out of a toaster oven. Same demonstrations and same policy architecture throughout, 25 trials per task. Fixed stereo camera: 27%. Wrist cameras added: 52%. Active gaze: 67%. When a grasped tool blocks the wrist cameras, that policy drops to roughly the level of the fixed camera. Gaze keeps working. The gaze is trained with RL, with no gaze demonstrations. Where it looks is where it reaches. When the gaze lands on the wrong object, the robot picks up the wrong object. That makes failures easy to read, and it suggests where to improve next. Paper: arxiv.org/html/2610.03710v1 Source: eyerobot2.github.io/ #PhysicalAI #RobotLearning #VLA

Uni-VLaT adds touch to pre-trained VLA policies for humanoid loco-manipulation, and the Back-Tap Walking result is the one that actually got my attention. Tap the robot's back, it walks. Stop tapping, it stops. Zero visual signal involved, that's the whole task. 0% success without touch, 85% with it, which isn't really a "touch helps" result, it's a "this task doesn't exist without touch" result. The rest of the numbers are more the usual incremental story, 68% to 75% average across five tasks from adding a prediction objective on top of raw tactile input, which is a real gain but not a dramatic one on its own. Also generalizes across GR00T and π0.5, so at least it's not a party trick tied to one backbone. Paper: ggkiller-air.github.io/Uni-V… Source: arxiv.org/html/2609.35450v1 Credit: @Tsinghua_Uni, @Beihang1952, @HKUniversity, @CUC1954_

Some robot tasks are basically impossible for a human to teleoperate cleanly. Not hard, impossible. Carrying a plate across a table with two independently controlled arms, for instance. Zero percent success rate, no matter the operator. GLIDE takes an interesting approach to this. An LLM reads the task, predicts where a human will fail, and writes a filter that sits between commands and robot to prevent it. Watches the data, catches what's still failing, rewrites itself. Same task, 0 to 70 percent. Two others go from 10 and 0 up to 90 and 90. Keep the guardrail running at deployment and a policy that barely worked alone jumps from 0 to 60-70 percent too. A good reminder that a lot of "the robot can't learn this" is actually "no human could demonstrate this in the first place." Source: guardrail-policy.github.io/

Most robots need hundreds of demonstrations to learn a task, then need hundreds more the moment anything about that task changes. This one needed one, and then just kept adapting. The system is VLBiMan++, out of @ShenzhenUni and @DexForceAI. Show it a single human demo of pouring water, and it breaks the task down into reusable pieces, then uses vision-language grounding to figure out how those pieces need to shift when the situation changes. No retraining, no extra demonstrations, no fine-tuning. The range of what "changes" means here is the part worth watching. It's not just moving a cup to a new spot on the table. The same one-shot prior carries across pouring, reorienting and unscrewing a cap, spoon and funnel tool use, zipping a pen bag shut, coiling a loose cable, folding a towel. Rigid objects, articulated ones, deformable ones, all handled by the same underlying approach, each needing its own way of tracking what "the object" even means as it changes shape. Then there's the embodiment jump. Skills learned on one dual-arm platform transfer directly to a completely different humanoid-style dual-arm robot, plugging in a pen, handing off an object, pouring, unscrewing, all without a single new demonstration on the new body. And to check it isn't just lucky on the first try, they ran the same task, uprighting a bottle knocked onto its side, twenty times in a row on the new platform and it held up. Every added axis here (new tasks, new objects, new scenes, new robots) is a dimension where most systems need fresh data. This one just needs to reason more carefully about what actually changed. Video: hnuzhy.github.io/projects/VL… Paper: arxiv.org/abs/2609.14310

A robot's camera predicts the future to decide what to do next. The problem is that prediction takes time, and time is the one thing a robot moving in the real world doesn't have to spare. That's the tension a team from @BAAIBeijing and collaborators went after with World Action Models, systems that use video generation to imagine what happens next before deciding how to act. The video model builds that prediction through several rounds of denoising, and each round adds latency to the control loop. The obvious fix is to cut it down to one round. The team found out why that quietly breaks things. Watching the denoising process closely, they noticed the background of a scene sharpens almost immediately. The gripper, the object it's about to touch, and how the two interact stay blurry until several steps later. Cut the process short and you keep a crisp picture of the room and lose exactly the part the robot needed to act on. Their fix, called DIDO, trains a single denoising step to hold onto that missing detail. A slower four step model acts as a teacher, and the one step model is explicitly guided to track the object, the gripper, and their interaction directly, with extra supervision pointed at exactly those regions rather than the whole frame. A pruning step then spends compute where the scene is actually changing and compresses the parts that aren't. The result is a model that predicts an action in 384 milliseconds instead of 562, while doing just as well or better on manipulation benchmarks; 99.0 percent on LIBERO, and the strongest score among methods with no embodied pretraining on RoboTwin. It also held up on a real Galbot G1 arm across four tasks, including with clutter and repositioned objects added in. It's a good reminder that not every shortcut is free. Cutting a process down to save time can end up cutting away the part that made it work in the first place, and the only way to know is to look closely enough to see what's actually happening at each step. Paper: arxiv.org/abs/2609.15570 Project page: loveju1y.github.io/DIDO/

Try holding your arm perfectly still for two minutes while someone else moves it into a sleeve. Now imagine doing that if holding still wasn't something your body could easily do in the first place. That's the gap in a lot of robot-assisted dressing research, which usually assumes a static arm. A paper currently under review takes this on directly. It fuses visual and force feedback into a single policy, trained first in simulation and then fine-tuned with real-world data, so the robot adapts to natural, uncooperative arm movement instead of requiring the person to stay put. It also handles long-sleeved garments, which add friction and occlusion that shorter sleeves don't. The results are the part that stands out. Tested across 12 participants and 264 real dressing trials, the system successfully dressed an average of 85 percent of each person's arm length, despite the arm moving throughout. This is the same principle behind a lot of the manipulation research worth watching right now. Fusing force and vision into one policy, rather than treating them as separate systems bolted together, tends to be what makes robots reliable outside a controlled lab setting. It's still under review, so treat the numbers as early rather than final, but the direction is one worth following. Paper: arxiv.org/html/2609.04759v1 Source: anonymous.4open.science/w/dr…

RT @Neuracore_AI: There's about $50-60 billion worth of electric motors in use in industrial robotics today. To make the more extreme predi…

Most of the robots sitting on factory floors right now are more capable than the tasks they're being used for. I've spent over a decade in robot learning, PhD at Imperial, postdoc at Berkeley, building the Dyson Robot Learning Lab, and the pattern hasn't changed. Every major step, from early imitation learning to today's VLAs and diffusion policies, was deployable years before most teams picked it up. The capability was never the bottleneck. The hardware in front of most companies right now can already do far more than it's doing. What's missing isn't a robot arm or a breakthrough model, it's the data and infrastructure to teach the hardware they already own new skills. That gap is what's costing teams the opportunity in front of them. Not a lack of technology, a lack of the layer that turns existing hardware into new capability. That's the layer @Neuracore_AI builds. Want to see what your hardware could actually be doing? Head to the comments to book in a demo with our team.

Teaching a quadruped robot to push an object it can't grasp is a hard RL exploration problem. The task reward stays at zero until contact happens, so a standard single-critic PPO ends up optimizing smoothness and energy penalties instead, and never finds contact at all. Members from @Unipisa, ETH Zürich & @nvidia address this with a separate exploration critic trained on a dense contact-seeking reward, guiding the end-effector toward candidate contact points, then decaying its weight over training so the policy shifts from contact-seeking to task-optimal once the physics has actually been found. Candidate points come from a general-purpose grasping algorithm, so the approach generalizes across object geometries without hand-tuning per task. Validated on a real quadrupedal mobile manipulator transporting chairs, transferring zero-shot to unseen IKEA furniture, recovering from failed contact attempts, and remaining stable under loads beyond the robot's rated capacity. A clean approach to a common problem: rather than hand-shaping one messy scalar reward, split exploration from task performance and let one decay into the other. Paper: tolomeis.github.io/contact-g… Project page: tolomeis.github.io/contact-g…

FANUC, @ABBRobotics and @Universal_Robot arms trained on @Neuracore_AI, running autonomously all day at our first event.

We just wrapped our first event at @Neuracore_AI! Thank you to everyone who turned up and kept the conversation going long after the last talk. Make sure to follow the Neuracore page to keep in the loop for more events like this, and check out the discussion over on our YouTube, coming later this week!