Users praised LingBot-Video for generating beautiful and realistic videos of humanoid robot actions.
Based on 2 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
Some beautiful and realistic videos created with LingBot-Video 🤗 Hugging Face: https://huggingface.co/collections/robbyant/lingbot-video https://x.com/rohanpaul_ai/status/2075313770830967154/video/1
🧵 2. 📄 Paper: https://github.com/Robbyant/lingbot-video/blob/master/paper.pdf The data story is the part that makes it feel less like a generic video model. Web video mostly teaches appearance. LingBot-Video adds 70,000+ hours of embodied footage: robot manipulation, navigation and egocentric video. That gives the model more action-and-consequence examples, not just pretty frames. The key part is Stage 2 (below image): that green video block is where LingBot-Video adds the 70,000+ hours of embodied footage, meaning robot manipulation, navigation, and first-person action videos enter the pretraining mix instead of relying only on normal web video.
🧵 9. LingBot-Video does not treat training data as a giant messy pile of images and clips. Every sample gets tagged across quality, camera, motion, semantic and structural signals, so the team can filter weak data, rebalance rare cases, and push more useful physical-action examples into training. The big deal is that the data mix is being engineered, not just scraped. That is how the model gets more exposure to motion, objects, actions and scene structure instead of only learning from whatever looks visually nice.
🧵 10. This figure shows how LingBot-Video organizes training data by what is in the scene and what action is happening, instead of treating every clip as just another random video. The big deal is the feedback loop: if some object or action category is rare, hard, or producing high loss, the system can up-weight those samples; if a category is already saturated or too easy, it can be down-weighted. So the model gets pushed toward better coverage of difficult physical actions, not just more of the same easy web-video patterns.
🧵 6. The scaling recipe is also so interesting They tried more routed experts under the same active compute budget, and 128 experts gave most of the gain without the extra overhead of 256. Fine-grained routing also beat coarser routing at the same 13B total scale https://x.com/rohanpaul_ai/status/2075313742787858649/photo/1
🧵 6. The scaling recipe is also so interesting They tried more routed experts under the same active compute budget, and 128 experts gave most of the gain without the extra overhead of 256. Fine-grained routing also beat coarser routing at the same 13B total scale https://x.com/rohanpaul_ai/status/2075313742787858649/photo/1
🧵 8. The bigger robotics angle is Action-to-Video. Give the model an initial world state and a robot action sequence, and it rolls out the likely future frames. That makes the video model closer to a world simulator: useful as a data engine, a policy evaluator and an action planner.
More results from LingBot-Video by @robbyant_brain 🐙 GitHub Code: https://github.com/Robbyant/lingbot-video https://x.com/rohanpaul_ai/status/2075314292493365574/video/1
More LingBot-Video results 🐙 GitHub Code: https://github.com/Robbyant/lingbot-video https://x.com/rohanpaul_ai/status/2075313775759278588/video/1
Some more fine creation from LingBot-Video https://x.com/rohanpaul_ai/status/2075314082849476948/video/1
Users praised LingBot-Video for generating beautiful and realistic videos of humanoid robot actions.
Based on 2 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
🧵 8. The bigger robotics angle is Action-to-Video. Give the model an initial world state and a robot action sequence, and it rolls out the likely future frames. That makes the video model closer to a world simulator: useful as a data engine, a policy evaluator and an action planner.