Ant Group Open-Sources LingBot-Video for Embodied AI

Ant Group's Robbyant has officially open-sourced LingBot-Video, a video foundation model designed for embodied AI. Rather than pursuing traditional visual aesthetics, the model focuses on understanding real physical laws, aiming to serve as a foundational component for a robotic 'brain' in applications like data engines, policy evaluation, and action planning.

Key Details and Architecture

LingBot-Video utilizes a fully customized DiT + MoE (Mixture of Experts) architecture. The model has a total parameter count of 30B but activates only 3B parameters during inference. This design significantly boosts inference efficiency over dense architectures when processing 1M long token sequences. The team claims it is about 3x faster on long inputs than dense models, making it suitable for real-time robotic planning. Additionally, the model supports interactive use, generating videos up to 1 hour long, and includes an agent to drive the process. It was pre-trained on large-scale internet video data and incorporated 70,000 hours of embodied data.

Training Mechanism and Evaluation

Changing the traditional scoring mechanism that only pursues visual beauty, the team employed a single-step GRPO algorithm with 6 precise reward signals: visual quality, text-image alignment, motion dynamics, motion coherence, human motion consistency, and physical rationality. This prevents the model from 'faking' physical effects. In the public RBench benchmark, LingBot-Video achieved an average score of 0.620, leading other open-source models in manipulation, long-horizon, and quadruped tasks, though the team admitted it still trails the closed-source Wan 2.6 in spatial and robotic arm tasks.

Open Source Availability

The model's code and weights have been fully open-sourced under the Apache 2.0 license. Official resources including model files, code repositories, research papers, and project pages have been released for developers to access and reproduce.

2026-07-09 ~ 2026-07-10 · 26 related posts

5 near-duplicate retellings: _akhaliq · thetripathi58 · dr_cintas · rohanpaul_ai · aigclink