Rajbalaji V
18 posts
Rajbalaji V
@rajships
Building Dexops | @fdotinc Canopy | Robot Learning & physical ai
BLR - SF Katılım Aralık 2025
124 Takip Edilen36 Takipçiler
RIC peaked 🚀
Met some cracked folks building physical AI in Bangalore. Great convos.
Thanks @spc_india, @RoboticswithV & HSR FC 🙌
Varun R@RoboticswithV
@spc_india hosted Robotics India Community and HSR Founders club for a buzzing evening over World Models and Physical AI in Bengaluru today! Insanely knowledgeable crowd and all of them are locked in to building Physical AI for the World, from India! We also celebrated 2000 memberships for Robotics India Community today with loads of Pizzas!🍕 Looking forward to elevating India’s Aura in frontier technologies! Next up : Hyderabad, stay tuned for the updates as we are bringing Physical AI to your city!
English
Called it in Paris.
When I was at #GOSIMParis2026 I got a preview of what @huggingface was building with the hashtag @LeRobotHF team and I immediately knew this was going to be a big deal. Today it's official: the LeRobot Humanoid is out.
Open source. 3D printed. ~$2,500 in parts. Full stack release: hardware files, assembly docs, simulation assets, training environments, real-world control. Just how the SO101s made arms accessible, this should make it possible for a whole new audience to build their own humanoid at home.
And I love that philosophy: this isn't trying to be the most advanced humanoid. It's trying to be the one you can actually build, break, repair, and learn from. For someone who routinely breaks stuff to learn, this is so very welcome!
Here's a little preview I filmed in Paris, and as I hinted - I'm absolutely planninbg to try assembling this on Back to Engineering. In pink of course. And maybe in a live stream 🤔
Can't wait to see this evolve!
English
@LeRobotHF The biggest thing here is making RL algorithms feel plug and play instead of tightly coupled to the training stack.
That’s probably more important long term than the SAC result itself
English
🤖Adding new RL algorithms to LeRobot just got much easier.
Demo: HIL-SERL training with a SAC-based RL algorithm on an SO-100 for a hole-in-hand peg-in-hole task.
Sparse reward, only 30 offline demos mixed with live robot experience, and ~1 hour of online training with human interventions only when the policy fails.
The bottom graph tracks intervention rate: high at the start, steadily dropping as the policy improves.
The refactor separates algorithm logic from training infrastructure:
• RLAlgorithm owns learning logic
• RLTrainer handles orchestration
• DataMixer combines rollouts, demos, interventions, and future data sources
Adding an RL algorithm now looks much closer to adding a policy: one algorithm file, one config, one registry entry.
SAC is first. RLT, RECAP, ConRFT, QC-FQL, DSRL, and VLA RL fine-tuning next!
@Thom_Wolf @ClementDelangue
English
@fdotinc We're here to put a dent in the universe. Otherwise why else even be here?
English
@andimarafioti This seems way more economically viable than realtime APIs, but how are you handling wake word / VAD / interruptions locally without the UX getting noticeably worse?
English
Reachy Mini just got a new Brain!
We released a fully open-source backend for talking to Reachy Mini. In the last 48 hours, 3,000+ robots have already hit the version deployed on Hugging Face's infra.
Until today, if you wanted a real-time voice agent in your robot, you were looking at $20+ a day. Realtime APIs are expensive because every second the connection is up, you are being charged! But with our approach, you can run the audio models locally and use your OpenAI or Claude subscription for the LLM, at no extra costs. The audio pipeline is very efficient; in the video, it's running on the cheapest Mac mini, 16GB of RAM (the one that sold out a couple of months back when people were buying them up for openclaws setups).
Tutorials coming in the next few weeks: how to run your robot fully locally, or with your favorite agent. Stay tuned!
English
Rajbalaji V retweetledi
China's Wuji Tech has confirmed that its highly dexterous Wuji Hands are featured in the French startup Genesis AI's new viral demo.
Paris-based Genesis AI, launched in 2025 with more than $100 million in seed funding, shared footage of a dual-arm manipulator robot autonomously performing a wide range of complex physical tasks with multiple steps using the same neural network. The robot in the demo is integrated with a pair of Wuji Hands that boast a reported 22 degrees of freedom (DoF) each.
“We believe GENE-26.5 is an early but important step toward human-level robotic manipulation,” the startup said in a blog post. “Manipulation capability does not emerge from model training alone. It requires a coherent system: hardware that can express rich contact, data collection that preserves human interaction, control that minimizes the gap between intention and execution, models that absorb multimodal supervision at scale, and evaluation infrastructure that makes iteration scalable and reproducible.”
The artificial intelligence is trained on data collected using high tech gloves that capture motion, force, and tactile feedback from real human hands at work. It’s combined with data from first person demonstration videos and large-scale simulation.
Genesis AI, which is led by former researchers from Mistrial AI and Carnegie Mellon University (CMU), says this helps the system better understand both the physical world and how humans interact with it. The startup says it’s developing a full stack robotics platform for use in factories, logistics hubs, laboratories, and eventually households.
English
Feels like we’re getting scarily close to human-level dexterity
Have VLAs + world models ever been connected this tightly to joint-level control before?
Feels like this could either massively reduce power waste or create a whole new compute bottleneck
@zhou_xian_ @ericschmidt
Genesis AI@gs_ai_
We are back. After one year of quiet building. Introducing GENE-26.5, our first robotic brain that takes a major step toward human-level capability. For years, robotics has struggled to learn from the world’s largest and valuable data source: Humans. Solving it means rethinking the whole stack from the ground up: - A robotics-native foundation model. - A 1:1 human-like robotic hand. - A noninvasive data collection glove for motion, force, and touch. - A simulator that turns weeks of experiments into minutes. GENE-26.5 is trained across language, vision, proprioception, tactile, and action. We designed a set of tasks to test how far we can go with this new paradigm. Fully autonomous, 1x speed, one model, same weights. (Enjoy with sound on) We are approaching the endgame for robotics. And this is just a beginning.
English
@hthieblot Agree on speed 💯
But speed compounds the most when paired with first-principles thinking, clear positioning, and obsessive distribution.
English
The best founders I know move at a speed that looks impossible to normal people. They compress 1 months in 3 days.
They ship 3 different ideas in a week.
Test 3 growth channels in one day.
Wake up the next morning and make 20 new videos with different hooks.
following up relentlessly
Operate 24/7, always talking to users and answering questions
Do hundreds of calls in a day.
Find and Schedule 200 investor meetings back-to-back for weeks.
Decide with 30% of the information and iterate/pivot while everyone else is still polishing their logo.
They are constantly in motion.
Most startups don’t die from bad ideas.
They die from waiting too long.
Speed creates momentum.
Momentum creates luck.
Luck creates opportunity.
When in doubt, move faster.
English
My very first video diving into #robotics has reached more than a quarter million views 🥹
So insane. I’m so grateful to everyone who watched it!
This is what made me go all in with #PhysicalAI 6 months ago - and it’s been the most amazing journey ever! 🤖💗
Now let’s see if we can top this 🔥☄️🌝
English
He discovered how to make any metal on the periodic table
And the only byproduct is water
Introducing Radify Metals:
English
Just dropped this nice little thing: mdshelf.
Turn any folder of Markdown files into a fast, beautiful, browsable site.
Point it at OpenClaw, Hermes, GBrain, wikis, notes, docs, whatever - mount each one under its own private URL and consume them anywhere.
English
@agazdecki You shouldn't sell just to sell. You need to work towards it,
1. You do your stuff, build the best thing
2. Stay discoverable
Customers will find their ways to buy
English
If you are a profitable startup, you have the only leverage that actually matters in an acquisition in 2026. You don't need to sell.
This means you can walk away from bad terms or lowball offers without hesitating.
Buyers can smell desperation but they also respect founders who are willing to keep building if the deal isn't right.
Use that leverage.
Don't just look at the top line price, look for the buyer who actually understands what you built.
PS - if you're looking to sell in 2026, we'd love to talk with you. Find time to speak with our team here: acquire.com/guided-by-acqu…
English
@TheGeorgePu Because we cannot see our own flaws as clearly as we see others', developing self-awareness and understanding common psychological traps is a crucial part in the self- leadership.
English
'I want to be my own boss.'
Cool. Have you met yourself as a boss?
Some people escape a bad manager just to become one.
No structure. No accountability. No deadlines unless a client sets them.
Freedom isn't the hard part. Discipline without external pressure is.
Most people need the cage more than they admit.
English