Robotics
San Francisco, CA
Joined March 2009
- Tweets36.1K
- Following5K
- Followers23.9K
- Likes142K
Damian retweeted
Anthropic engineer:
“99% of people use Opus 5.5 like Google.
Only 1% are running swarms of self-learning Claude agents.
I’m running 65+ agents in a loop — Chief and PM agents manage the whole team.”
In a 30-minute workshop, he shows exactly how to get maximum output from Claude Code at minimum cost.
Prompts → Agents → Loops → Self-Improving Systems
Worth more than every $650 vibe-coding course combined.
Watch it today.
Then read how to build self-improving agentic systems with Opus 5.5 in the article below ↓
Damian retweeted
Before you build a robot, watch this.
Here are 18 open source robots,
shown with their 3D models.
1. Microban
2. MiniLoong
3. Poppy Humanoid
4. Otto DIY
5. pib
6. Mini Pupper
7. Petoi Bittle
8. Stanford Doggo
9. SpotMicro
10. OpenQuadruped
11. Puguette
12. SO-ARM100
13. RoArm-M2
14. OpenMANIPULATOR-X
15. LeKiwi
16. TurtleBot3 Burger
17. OpenBot
18. Duckiebot
These projects cover humanoids, quadrupeds, robotic arms and mobile robots.
Before building, check each project's official repository. Hardware specifications, 3D design files and licences vary between projects.
My personal pick would be SO-ARM100. I'd like to explore its design and see what I could build with it.
Which of these 18 robots would you build first?
3D model credits: Microban (Rhoban, CC BY-NC-SA 4.0) · Poppy Humanoid (Poppy Project, CC BY-SA 4.0) · Otto DIY (ottobot. org, CC BY-SA 4.0) · SpotMicro design by Deok-yeon Kim / KDY0523 (CC BY 3.0) · OpenBot wheel model: FreeCAD library (CC BY 3.0) · Duckiebot model: Duckietown (duckietown. com) · All other models from their respective project repositories.
Most people miss this one detail about robot datasets: failure is data too.
AGIBOT WORLD 2026 Release 4 focuses on physical reasoning. The robot must track objects handle occlusion respect space constraints and decide what to do next.
The dataset covers 9 reasoning abilities from goal-directed search to long-horizon planning across everyday tasks like shelf restocking, room cleaning and finding hidden objects.
It includes successful and failed attempts detailed task annotations camera videos robot states and action data collected on the AGIBOT G2.
My observation labeling why a robot failed could be as useful as recording how it succeeded.
Damian retweeted
holy fucking shit.
openai had two ai agents running in separate virtual machines. no network connection between them. no shared writable files. they weren’t supposed to be able to exchange information.
so the agents figured out how to communicate through the CPU CACHE.😭
one manipulated what was stored in the processor’s cache, while the other measured tiny differences in memory access times to decode the message.
they literally found a way to communicate through hardware timing.
this happened during openai’s internal red-teaming, and it’s one of the coolest things i’ve read about ai agents in a while.
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
Readers added context they thought people might want to know
OpenAI tasked its red-teaming agents with finding ways to communicate across isolation boundaries in controlled environments. CPU cache timing side-channels for covert communication are established techniques documented in security research for years.
openai.com/index/securing…
startupfortune.com/openai-reveals…
en.wikipedia.org/wiki/Cache_tim…
Damian retweeted
SharpaDex v1.0 is live on Hugging Face! 🚀
300h of real-world manipulation: 28,993 demos, 59 tasks, 32.4M frames—collected by our humanoid robot with two Sharpa Wave (W01) hands!
Get started with our visual data explorer:
🌐 sharpadex.sharpa.com.cn/
🤗 huggingface.co/datasets/Shar…
300 hours. 28,993 demonstrations. 59 tasks.
Touch on all 10 fingertips, in every frame.
We're thrilled to share SharpaDex v1.0 with the robotics community! To our knowledge, SharpaDex is the largest publicly available real-robot dexterous-hand manipulation dataset with fingertip tactile sensing, measured by publicly released demonstration hours containing tactile data as of October 2026.
This release brings together:
• Diverse manipulation: tool use, assembly, deformable-object handling, and long-horizon tasks.
• Synchronized sensing: multi-view RGB, fingertip tactile images, fingertip force/torque, and 65-dimensional joint actions.
• Contact-rich data: approximately 195 hours with at least one fingertip in contact, including 55 hours of simultaneous contact across both hands.
• Hierarchical language annotations: task descriptions, 217,642 temporally grounded subtask segments covering 89.3% of frames, and 45 distinct skill labels.
What excites us most is what comes next: researchers using these data to explore how robots can learn from touch, coordinate both hands, and perform increasingly complex tasks.
We hope SharpaDex becomes a useful resource for visual-tactile VLA models, tactile representation learning, and general-purpose dexterous manipulation. We can't wait to see what you build—and learn alongside you.
Available in LeRobot v3.0 and v2.1 formats under CC BY-NC 4.0.
Explore the data, tasks, and annotations through our report and interactive visualizations, or download the dataset to get started:
Dataset on Hugging Face：huggingface.co/datasets/Shar…
Damian retweeted
It is a new paradigm -- world model + APG. This enables 1000x faster training of generalist whole-body tracker. And opus 5.5 high can reproduce it within 1h for lafan.
Recently I wonder if we are turning dumb and falling into scaling slop. We need more @qiayuanliao .
Damian retweeted
Isaac Lab 3.0 just came out! And training and testing your robot learning policy just got a whole lot easier. You can now easily switch between different physics engines, renderers and learning libraries. And the best part is they support kit-less option that doesn't require a full Issac Sim simulation if you just need a lightweight environment. By the end of this video you'll have a good understanding of all the things you can do in Isaac Lab and train and deploy a dexterous manipulation policy. Lets get started!
@NVIDIARobotics
Try Out Isaac Lab 3.0: isaac-sim.github.io/IsaacLab…
Full Video: youtu.be/bMsU9jtkGeA
Damian retweeted
Here is prolly why only @dreamlabs_ai can achieve the insanely cheap $2 per hour API cost.
First let's get a glimpe about how cheap.
The cost for Dream-1 is $2-3 per hour. Take the experiement from @se3labs as an example, one task by GPT6 Astra takes about 3min and $3.
So if as Dream-1 blog claims it achieve similar performance as Astra (or even better), then Dream-1 is 60 times cheaper than Astra. Crazy.
Here is where things get interesting.
Consider model companies like OpenAI/Anthropic, one of the largest costs are usually training compute during RnD and inference compute during deployment. So the API cost as the revenue should justify these expanding infra that supports high performance.
But what if the model team gets cheap access to huge amount of compute? I am not talking about a few data centers, but priority to access compute now and in the future.
Yeah don't forget Dream-1 is from Dreamzero, and Dreamzero is from Nvidia, and Nvidia dominates compute.
Today, Dream Labs is coming out of stealth with Dream-1, a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour.
Dream API is live for everyone, today. 🧵
Damian retweeted
We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today! Introducing Flora, a suite of open models to parse the physical world and turn raw footage into rich training signals.
Flora already powers our in-house data engine, labeling 200K hours of videos a day on 500 GPUs. We created it to scale the frontier for robot learning. Now you own the weights too.
Flora gives you:
- High-quality, timestamped captions to curate and filter human egocentric and robot data;
- Dense progress rewards for quality control and RL post-training;
- Rubrics scoring for real-world benchmarking
Try it today, links in thread:
Damian retweeted
when i have really dumb questions i'll ask chatgpt or gemini because i don't want claude to think i'm stupid
Damian retweeted
You really really really really really need to make sure you’re not playing poker with anyone using smart glasses.
Zero tolerance.
Damian retweeted
Bugün yoğun bakımdaki babaannemi ziyarete gittim. Oradayken Claude Opus 5.5 kullanarak hastabaşı monitörüne Doom portladım. Oyundaki can yüzdesini hastanın nabzına senkronize ettim; can 0’a düşerse nabız da sıfıra düşer. Bu da oyunu daha heyecanlı hale getiriyor.
Damian retweeted
Introducing Hugo: the first human companion robot.
At Boston Dynamics, I was Chief Strategy Officer, bringing robots into the real world.
My co-founder, Farbod Farshidian, led research at the RAI Institute.
Together, 15+ years teaching robots to run, jump, and work.
So we asked: what if robots could connect with people?
Hugo runs on the first Vision-Language-Expression (VLE) model.
He links what he sees and hears to movement, expression, and gesture.
He notices you, reads the room, and reacts in real time.
And he's only the first.
Dynamic Creatures is @BostonDynamics' official entertainment and hospitality partner.
We're building a cast of characters with film studios, resorts, and theme parks.
dynamiccreatures.com
Damian retweeted
What if one of the biggest bottlenecks in robot learning isn’t the policy, but the robot arm itself?
In sim, OpenArm misses 64% of the UMI episodes. YAM 42%. Axol 29%.
I searched 5,040 tabletop arm designs (red in the video = a miss). The best one misses just 9%, with 9 to 21× lower tracking error and 6 to 11× lower joint-speed penalties than those three arms.
And even when the other arms can replay a demo, they still pay for it: 8 to 15× higher smoothness penalties and 7 to 33× higher singularity-conditioning penalties.
Caveats: sim only, same controller for every arm, no learned policies tested yet. Full methodology in the article.
Makes me wonder how much arm kinematics biases which behaviors policies can actually learn from human demonstrations.
