LingBot-VA/VLA 2.0 Released: Native Embodied Foundation Model
An Ant Group-backed embodied AI company (@robbyant_brain) released the LingBot-VA 2.0 and LingBot-VLA 2.0 models. Instead of the common approach of bolting an action head onto a video generation model, LingBot is trained from scratch as a native video-action foundation model to solve general robot control and cross-embodiment generalization.
Core Design and Technical Details
Authors @omarsar0 and @heyshrutimishra pointed out that pure video generation models only learn appearance, not how actions change the physical world. LingBot places world states and latent actions in the same semantic latent space, utilizing self-supervised learning on unlabeled web videos. Architecturally, it uses a 30B MoE design (activating only 3B parameters during generation) with a full causal Transformer. To achieve real-time control, system optimizations such as dual-expert distillation, FP8 TensorRT, and paged KV-cache were applied, reducing single-block latency from 927ms to 142ms. Furthermore, @omarsar0 highlighted a "closed-loop" inference mechanism that drafts the next action chunk based on the latest observations to avoid lookahead drift.
Evaluation and Generalization
In evaluations, the model achieved an average score of 93.6 on RoboTwin 2.0, surpassing models like π0.5, with stable performance across clean and randomized scenarios. @HeyToha noted that the model also scored 0.620 on the RBench benchmark, leading overall in tasks like long-horizon manipulation and quadruped control. The LingBot-VLA 2.0 version translates different robot controls into a unified 55-dimensional action language, successfully covering 20 robot configurations—including arms, dexterous hands, and bases—and can adapt to new environments with just 10–15 demonstrations.
2026-07-10 ~ 2026-07-12 · 24 related posts
- Embodied Native Robot Model Released — heyshrutimishra · 2026-07-10
- LingBot-VA2.0 Focuses on Embodied Native Pretraining — 机器之心 · 2026-07-10
- LingBot-VA2.0: Embodied Native VA Model — 量子位 · 2026-07-10
- LingBot-VA 2.0 Released — chris_j_paxton · 2026-07-10
- LingBot-VA 2.0 Released — CyberRobooo · 2026-07-10
- Robot Control Shifts from Reactive to Predictive — Purple-Low-2779 · 2026-07-11
- Open-Sourced LingBot-VA 2.0 Robot Foundation Model — philfung · 2026-07-11
- Inside the LingBot-VA 2.0 Robotic Foundation Model — Aiden_Tech_Ai · 2026-07-11
- LingBot-VLA 2.0 Robot Control Model — LearnWithBishal · 2026-07-11
- [source] LingBot-VA 2.0 Upgrades Robot Control Stack — omarsar0 · 2026-07-11
- Video-Action Foundation Model for Robot Control — omarsar0 · 2026-07-11
- Video-Action Embodied Model Trained From Scratch — omarsar0 · 2026-07-11
- Real-Time Optimization Details for MoE Embodied Video Streaming — omarsar0 · 2026-07-11
- Closed-Loop Inference Enables Real-Time Robot Actions — omarsar0 · 2026-07-11
- Embodied Model Leads on RoboTwin 2.0 — omarsar0 · 2026-07-11
- [source] Ant Group's Embodied Video-Action Model Tops Benchmarks — omarsar0 · 2026-07-11
- 30B MoE Embodied Model Activates 3B Parameters — HeyToha · 2026-07-12
- Embodied Video Models Prioritize Physical Consistency — HeyToha · 2026-07-12
- Embodied Video Model Takes the Lead on RBench — HeyToha · 2026-07-12
- World Model Project Tailored for Robotics — HeyToha · 2026-07-12
- [source] LingBot-VLA 2.0 Whole-Body Robot Policy — rohanpaul_ai · 2026-07-12
- LingBot-VLA 2.0: Universal VLA Model for 20 Robot Morphologies — rohanpaul_ai · 2026-07-12