Teaching Robot Actions with Unlabeled Video
rohanpaul_ai · x · 2026-07-14
This covers the representation learning method of LingBot-VA 2.0:
- Standard video tokenizers typically only learn to "reconstruct pixels".
- This approach aligns compressed states with the semantic features of a frozen vision model, while using inverse dynamics and forward dynamics to learn compact latent actions from frame-to-frame changes.
- Consequently, unlabeled web videos can supply action-relevant training signals without relying on robot action annotations.
Related event: LingBot-VA 2.0: A Control-Native Foundation Model for Robotics(9 posts)→
More from Embodied
- Orlando robotaxi ride goes unsupervised in a Model Y — aelluswamy · 2026-07-21
- Anthropic and Physical Intelligence held early M&A talks this spring — steph_palazzolo · 2026-07-21
- EPO improves 3D foundation models by aligning edges, poses, and depth — ducha_aiki · 2026-07-21
- Applied Intuition launches Dana for physical AI, alongside a16z interview on one billion machines — a16z · 2026-07-21
- Hearing Aids Embrace AI: Auracast and OmegaAI Reshape Personal Audio Broadcasting — BrandonSawalich · 2026-07-21
- a16z Talks with Applied Intuition: The Next Decade of Physical AI — a16z Podcast · 2026-07-21