正在加载视频...
视频加载失败
🤖 LingBot-VLA 2.0 is now open-source — our next-gen embodied foundation model. 🔷 60,000 hours of high-quality pretraining data — combining curated robotic demonstrations and egocentric human operation videos 🔷 20 robot configurations across 17 brands — Astribot, Leju, Unitree, Franka, Fourier, Realman, and more 🔷 Whole-body DoF: heads,... show more
35 条评论

📈 GM-100 Bimanual Benchmark (Generalist Setting — all models evaluated as generalists, no task-specific fine-tuning): 🦾 AgileX Cobot Magic (Progress Score / Success Rate): • LingBot-VLA 2.0: 66.2 / 34.4% • π0.5: 59.1 / 32.2% • GR00T N1.7: 36.3 / 17.8% 🦾 Galaxea R1 Pro: • LingBot-VLA 2.0: 34.6 / 15.6% • π0.5: 27.4 / 8.9% • GR00T N1.7: 16.4 / 5.6% On long-horizon mobile manipulation, LingBot-VLA 2.0 consistently outperforms π0.5 across both in-domain and out-of-domain settings — demonstrating stronger cross-embodiment mobile manipulation capability.

Beyond benchmarks — deployment is underway. ⚡️ Inference under 130ms on RTX 4090. We're working with ecosystem partners on commercial pilots in retail sorting, logistics, and manufacturing. Chip partners Horizon Robotics (S600) and NVIDIA (Jetson Thor & Orin) have completed model adaptation. Fully open-source. Developer events & toolkits coming soon. 🔗 GitHub: 🔗 HuggingFace: 🔗 Tech Report: 🔗 Website:

Next-level! 🤖

The expanded action space is honestly the biggest upgrade I've spotted

Love seeing more focus on practical deployment instead of just bigger models.

The whole body action space is the update that stood out most to me.

The engineering behind this project is remarkable. Extensive pre-training, broad robot compatibility, and low latency make it a strong foundation for the next generation of embodied AI.

Which task category has shown the greatest improvement in reliability between V1 and V2.0?

Congrats on open-sourcing LingBot-VLA 2.0! 🔥 The jump to 60k hours of data and whole-body control across so many platforms is impressive. Those GM-100 results look strong – especially the mobile manipulation gains. Any plans to share the training code or fine-tuning examples soon? Would love to try it on a Unitree setup.

Really like that the focus here is practical deployment over flashy demos. The expanded action space and open-sourced post-training code make it much easier for developers to actually build on this.

This is an impressive benchmark for embodied AI.

Really like the emphasis on making post training more accessible. That can have a big impact for developers building on open source models.

Making LingBot-VLA 2.0 open-source is the part I’ll be watching closely. The most interesting results may come from what developers discover once they start testing it on setups the original team didn’t design around.

Twenty robot configurations in one foundation model is honestly a bigger engineering challenge than it sounds.

Better data pipelines usually matter more than people give them credit for.

This is a verified based robot you will like to try out. All you need to do is study it and then you are fit to go

Wow, this release is really impressive! It's awesome to see such practical innovation happening in embodied AI.

More practical releases like this, please.

El nivel de desarrollo detrás de este proyecto es realmente impresionante. Miles de horas de entrenamiento, amplia compatibilidad con robots y una latencia muy baja forman una combinación muy sólida.

This is the inflection point for real-world commercial robotics. Nice!

The mix of open-source and commercial pilots makes this especially interesting.

This is a valuable contribution to the AI ecosystem.

Practical deployment always changes the conversation

Impressive release, great to see practical innovation in embodied AI.

Supporting 20 robot configurations in a single VLA model is an ambitious milestone. Looking forward to seeing how well it generalizes across different hardware.

The scale of the pretraining data is remarkable.

Filtering and curating 60,000 hours of training data is probably a bigger part of this release than most people will notice. For LingBot-VLA 2.0, I’d be interested to see which data-quality decisions had the biggest impact.

This feels like the kind of release that’s more exciting for robotics developers than headline readers.

Open-source robotics reaching new heights with innovation.

Generalist evaluation feels like a much more realistic benchmark for robotics.

60,000 hours of pretraining across 20 robot configs is a serious step toward general purpose embodied AI.

This feels bigger than another model release. Building an open foundation for developers is the kind of progress the ecosystem needs.

@xetgepete Curious to see how it performs across all of those platforms.

Curious to see how LingBot VLA 2.0 performs across such a wide range of robot configurations.

This is the kind of open research that pushes embodied AI forward. Great work!
