kewang

[Isaac Sin @IsaacSin12]

ONLY 8 hours away!

MakerMods Physical AI × OpenClaw Hackathon Shenzhen, March 28–29

350 hackers. 25 real robots. live VLA training.

6 tracks: hardware mods, sim-to-real, OpenClaw, industrial manipulation, and using our MakerMods App. all on XLeRobot.

with amazing sponsors @AMD @huggingface @LeRobotHF @QualiaRobotics @tnkrdotai @norma_core_dev @Neuracore_AI

couldn’t be done without you guys @Ryan_Resolution @GZinMetaverse @Robo_Tuo @QILIU9203 and 杨志超

this is what Physical AI looks like when it’s not on a whiteboard 🦞

[kewang @Alchemist1017]

I have been working on Pi05 to unleash its full power, namely subtask generation and knowledge insulation, and here it is: https://t.co/Ho7PZaKu7g. I also wrote a blog explaining the architecture and implementation details of pi05: https://t.co/cQSsYOtRER. Next: real robots!

[kewang @Alchemist1017]

@VilleKuosmanen I use OpenAI codex, it's not expensive as Claude Code and does pretty job, especially for debugging. People complaining about the slow speed, now with codex 5.3 it becomes much faster. I love CC but the token limit is low.

[Neuracore @Neuracore_AI]

𝗬𝗼𝘂 𝗱𝗼𝗻'𝘁 𝗻𝗲𝗲𝗱 𝗮 $𝟱𝟬,𝟬𝟬𝟬 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗿𝗼𝗯𝗼𝘁 𝘁𝗼 𝘁𝗿𝗮𝗶𝗻 𝗔𝗜 𝗽𝗼𝗹𝗶𝗰𝗶𝗲𝘀.

Our ML team is integrating the @CobotMy mechArm 270 Pi with the Neuracore platform.

The biggest barrier to learning robot learning isn't talent or ideas, it's access to expensive hardware. We've already open sourced Neuracore for academia. Now we're making it work with affordable hardware to remove that barrier completely.

The mechArm 270 Pi integration is coming soon. If you're a hobbyist, researcher, or educator working with accessible robotics platforms, stay tuned.

Robot learning is for everyone. We're building the infrastructure to prove it.

[kewang @Alchemist1017]

@VilleKuosmanen 1 hour if possible:)

[kewang @Alchemist1017]

@VilleKuosmanen Do you mean bfloats gets much higher loss?

[kewang @Alchemist1017]

@VilleKuosmanen It is already happening in china🤣

[kewang @Alchemist1017]

@ihorbeaver This can be solved with asynchronous inference, in which you start doing inference in the middle of action chunk, half of the chunk will overlap but will be no delay

[kewang @Alchemist1017]

@RemiCadene @pollenrobotics @rerundotio Is the VR teleoperation code open sourced as well? :)

[kewang @Alchemist1017]

@RemiCadene @pollenrobotics @rerundotio Very impressive! Just wondering how the demons are made? Through VR or other methods?

[kewang @Alchemist1017]

@GoddenThomas Maybe you can try to imitate an animation and might work better

[kewang @Alchemist1017]

@GoddenThomas What are you doing now?

[kewang @Alchemist1017]

@GoddenThomas is it possible to learn a slide actuator model? Or just a simple regression.

[kewang @Alchemist1017]

@GoddenThomas Do you improve the mechanical side of SLIDER or just doing better parameter estimation? I know the mechanical side is not very good

[kewang @Alchemist1017]

@GoddenThomas Impressive!

[kewang @Alchemist1017]

@GoddenThomas Great work! I also think the contact settings of stairs are problematic. Maybe the ball sensor under foot cause some issues as well. Furthermore, you can penalize contact forces as the foot impact to ground seems large