Rohan Seelan

@robertorobotics This is awesome! How many episodes were used to train the base policy?

With my custom architecture I now achieve on average 60% progress on my gearbox assembly task. The original average progress of pi0.5 base (after 10k fine tuning) was 14% . Trajectory of progress is looking good. Now have to reduce the cycle time of iterating on a base policy.

@PrathamJainAI Could I please get added?

@MattHartman @pravsels Dm’d

If you're a robotics company building on top of VLAs / other end-to-end learned AI models and want to train models reliable enough to deploy, ping me.

@davefontenot @hf0 Resonant.. applied

Your moat is your soul.

@SourishJasti Deploying robots on the East Coast, DM’d

If you’re deploying robots in the real world, let’s get in touch

@Noriaki_Hirose @CatGlossop @shahdhruv_ @svlevine Wow

Robotic foundation models generalize well—but high inference latency limits real-time deployment. 🚀 AsyncVLA enables real-time control of large robotic models, even under network delays. Great collaboration with @CatGlossop ,@shahdhruv_ & @svlevine ! #EmbodiedAI #EdgeAI

@jiazhi_yang2024 Ah I see, makes sense! Thank you for the clarification

Thank you for your interest! Yes, one rollout stage only. One major reason for the moderate recap performance is the policy pretraining. In Pi06*, recap method is applied in both policy pretraining and finetuning, upon a proprietary Pi06 architecture, whereas we can only re-implement it on a public open-pi05 checkpoint, which is not pretrained with action conditioning as additional input.

🧐Applying world models to improve real-world policy on challenging manipulation tasks used to be considered out of reach. 😌After sustained effort, we’re now seeing encouraging progress. 🚀Thrilled to introduce RISE: Self-Improving Robot Policy with Compositional World Model opendrivelab.com/kai0-rl/ arxiv.org/abs/2602.11075 RISE is, to our knowledge, the first work to use a world model as an effective learning environment for challenging real-world manipulation, enabling policy improvement on tasks that demand high dynamics, dexterity, and precision. Incredible teamwork with @lin_kunyang111 @francislee2020 @YueXiangyu @HaoZhao_AIRSUN @smch_1127

@LeRobotHF think this isn't getting the hype it deserves

A new reward model called SARM —> Stage-Aware Reward Modeling for long-horizon robot manipulation is now in LeRobot! 🤗 Long-horizon, contact-rich manipulation (think folding a T-shirt) is messy. Demonstrations naturally include hesitations, corrections, and variable quality. Classic Behavior Cloning (BC) treats every frame equally, SARM takes a smarter approach. 🔎 How does it work? SARM uses a video-based reward model to predict: • the current task stage • fine-grained progress within that stage (0 → 1) This enables Reward-Aligned Behavior Cloning (RA-BC) by reweighing data based on progress improvement, allowing the model to distinguish trajectories that make progress from those that stall. Left image: successful rollout, learned progress smoothly increasing 0 → 1 Right image: unsuccessful rollout

@babugi28 @shloke_patel @ycombinator Congrats guys

My mom just sent me this video of my co-founder @shloke_patel and me pitching our mango business when we were 13. Now, in 50 days, we’ll be pitching at @ycombinator's Demo Day in SF. None of this would've been possible without our rockstar parents. Order some mangoes this summer at mangounited.com

@jackvial89 gotcha, thanks!

@RohanSeelan thank you, I’d recommend checking out the RTC implementation in LeRobot, that will get you a lot of the way there. I do plan to make my improvements open source soon

60hz! real time chunking on an so101 with LeRobot. Not looking too bad. Bit of jitter throughout but no mode switching across chunks

@BrianRoemmele Company name?

“Go fold those towels. Snap, snap”

@DynaRobotics @JasonMa2020 @HumanoidsSummit Is there a recording of this?

DYNA's Cofounder & Chief Scientist @JasonMa2020 will be speaking at the @HumanoidsSummit on Dec 11 &12! Jason will break down how @DynaRobotics is building robust robot foundation models for real-world manipulation and what it takes to scale.🦾 From research labs to production deployment—don't miss this! Come say hi if you are around!  #HumanoidsSummit #Robot #AI

@Deeplearner2 @7Xrobotics Super cool? Is this a fine tuned open source model?

Our robot just got an upgrade. The eyes (camera), arms, and brain (controller) are now one compact portable unit — easy to move to any table, no perfect alignment needed, and fewer wires @7Xrobotics

@jpizarrom @LeRobotHF Amazing work!

🤖 early-stage experiment finetuning SmolVLA + RECAP-style advantage signals (inspired on π*0.6 paper) as BC actor + one-step flow actor with AWR on SO100 @LeRobotHF linkedin.com/feed/update/ur… previous experiments ACFQL + HIL-SERL x.com/jpizarrom/stat… x.com/jpizarrom/stat…

@LeRobotHF @RemiCadene Amazing work! How many episodes were used to create the cloth folding checkpoint?

We’re releasing: • 6 X-VLA checkpoints, including a pretrained base model huggingface.co/collections/le… • A cloth-folding checkpoint achieving 100% success for two straight hours • A freshly ported folding dataset • Full docs, blog, and paper 📄 huggingface.co/docs/lerobot/e… A small preview of something bigger coming soon 👕👀

🚀 Introducing X-VLA ; LeRobot’s new soft-prompted Vision-Language-Action model. X-VLA is built to scale across many embodiments: different robots, cameras, action spaces, and environments, all handled by one unified transformer backbone. - Generalist across robots (Franka, WidowX, Agibot, sim + real) - Soft-prompt domain IDs let the model adapt to new hardware with tiny learnable embeddings - Flow-matching + transformer core for smooth, continuous 50 Hz control - Pretrained on a mixed-embodiment dataset spanning 7+ platforms and diverse tasks - Fine-tune on any dataset using one of the 6 checkpoints we provide out of the box.

@VilleKuosmanen Some sort of value function?

testing something new

@elonmusk Please approve my candidate packet for my Tesla co-op position! I interviewed for the controls engineering co-op position at Austin. It’s for the Model Y manufacturing line. I was told the executive team needs to approve it before any potential offer is discussed!

Congrats to the advancing teams in the Not-a-Boring Tunnel Competition! Started with 390 competitors, down to the 12 finalists, aka The Digging Dozen. boringcompany.com/competition