Open-Sourced Embodied Video Model LingBot-Video

HeyNayeem · x · 2026-07-12

Robbyant has open-sourced LingBot-Video, emphasizing that it is designed specifically as a foundation model for embodied intelligence (a "robot brain"), rather than for general-purpose video generation.

Key highlights:

Evaluations for action-conditioned generation/rollout show it generalizes to unseen data rather than just memorizing training trajectories.
The post-training phase utilizes 6 types of reward signals: aesthetic score, visual quality, image-text alignment, dynamics, motion consistency, human motion consistency, and a specialized physical feasibility signal.
Data-wise, beyond internet video, it incorporates over 70,000 hours of embodied data spanning robotic manipulation, navigation, and egocentric content, trained via a 5-stage curriculum from 192p → 1080p.
Architecture is Sparse MoE: 30B total parameters with only 3B activated per step. The author notes that on 1M token sequences, throughput reaches roughly 3x that of an equivalently sized Dense 30B model.
On evaluation, it scores an average of 0.620 on RBench, leading both open and closed-source baselines in the table. However, the closed-source Wan 2.6 still outperforms in spatial, single-arm, and dual-arm tasks, though it doesn't dominate universally.

Original post →

Open-Sourced Embodied Video Model LingBot-Video

More from Embodied