> 2026-07-08/09 · Ant Group's Robbyant (Lingbo Technology) · Embodied AI Infrastructure > LingBot-VLA 2.0: https://technology.robbyant.com/lingbot-vla-v2 > LingBot-Video: https://www.ithome.com/0/974/517.htm > LingBot-World 2.0: https://www.ithome.com/0/974/508.htm > Hugging Face: https://huggingface.co/collections/robbyant/lingbot-vla-v2
The Event: One Team, Three Open-Source Models, a Full Stack
On July 8–9, 2026, Robbyant, an Ant Group subsidiary, open-sourced three embodied-AI foundation models in a single stroke:
| Model | Params | Positioning | Key benchmarks | |------|--------|------|----------| | LingBot-VLA 2.0 | 6B | Vision-Language-Action model, directly outputs robot actions | Leads π0.5 on GM-100 bimanual tasks | | LingBot-Video | 30B (3B activated) | Video foundation model for embodied AI | RBench 0.620; first on Physics-IQ Verified | | LingBot-World 2.0 | 14B | Real-time interactive world model | Stable 720p/60fps output for 1 hour |
Together they form an embodied-AI "VLA + video foundation model + world model" trio — positioned against NVIDIA Cosmos, but on a fully open-source, Apache-2.0 path.
The engineering significance lies not in any single model's strength, but in three same-lineage models sharing one stack and being open-sourced together.
The Three Models in Depth
LingBot-VLA 2.0: 6B params, unified action representation across 20 robot embodiments
VLA (Vision-Language-Action) models map camera images plus language instructions to robot actions. Prior VLAs had two chronic problems: each robot embodiment required separate training (different joint definitions), and slow inference at deployment (most exceed 200ms).
LingBot-VLA 2.0's solutions:
1. Unified action representation — a 55-dimensional vector unifies states and actions across embodiments:
- Arm joint positions: 14
- End-effector poses: 14 (XYZ + quaternion × 2)
- Gripper positions: 2
- Finger joint positions: 12
- Waist: 4
- Head: 2
- Mobile base signals: 3
- Reserved: 4
- 50,000 hours of robot trajectories (20 robot configurations, from single-arm to full humanoid)
- 10,000 hours of egocentric human video
- Rich character actions (casting, attacking, jumping)
- Text-driven events
- Built-in Pilot Agent + Director Agent for continuous world evolution
- Simultaneous multi-user interaction
- Stable 720p/60fps real-time output for up to 1 hour without quality decay
- Data mutually reinforces: the VLA's 50,000 hours of robot trajectories teach the video model physically consistent robot motion; the video model's synthetic data can expand VLA training sets.
- The world model provides simulation for the VLA: closed-loop training in simulated new environments before deployment — an engineering compromise where real-world RL is infeasible.
- Shared pretraining backbone: Qwen3-VL-4B as the VLA's vision-language backbone, the same ViT for the video model, and the world model's Pilot Agent reusing VLM instruction understanding.
- Third-order jerk (motion smoothness)
- Velocity/acceleration Z-scores
- URDF reconstruction vs. original-state video consistency
- MANO-reconstructed hand poses
- 07-08: Nature paper (Unitree G1 in-vivo surgery)
- 07-08: LingBot-VLA 2.0 open-sourced
- 07-09: LingBot-Video open-sourced
- 07-09: LingBot-World 2.0 open-sourced
Robots lacking a component simply pad zeros. One model can therefore control single-arm, bimanual, humanoid, and mobile robots.
2. MoE action experts — transformer feed-forward layers are replaced with sparse MoE (one shared expert + routed experts, top-K per token). At equal activated parameters, MoE shows lower training loss and validation action error than dense baselines.
3. Data-quality engineering — 60,000 hours of pretraining data:
The filtering pipeline explicitly measures third-order jerk plus velocity/acceleration Z-scores, filtering out overly static or jitter-anomalous samples. Video is validated against original states via URDF reconstruction; human-hand video uses MANO hand-pose reconstruction.
4. Dual-query distillation — two learnable queries over visual/text tokens: Q_t attends to the current observation, Q_{t+T} to future observations, guided by two teachers (LingBot-Depth for geometric supervision, DINO-Video for temporal dynamics).
Inference speed: 130ms on an RTX 4090D (10 denoising steps) — practical speed on consumer GPUs.
Key benchmark: on GM-100 (100 bimanual tasks) with AgileX Cobot Magic, LingBot-VLA 2.0 scores 66.2/34.4 (progress/success), beating GR00T N1.7, π0.5, and LingBot-VLA 1.0. On long-horizon mobile manipulation (Astibot S1 fridge sorting, Cobot Magic–ARX X5 stove cleaning) it beats π0.5 in both in-domain and OOD settings.
LingBot-Video: 30B total / 3B activated MoE video foundation model for robotics
Video generation has two routes: toward cinema (creative content) or toward robots (physical-world understanding). LingBot-Video takes the second.
Architecture: DiT + MoE. 30B total parameters, only 3B activated at inference — roughly 3× more efficient than a dense model of equal scale.
Data: a data-profiling engine adds 70,000 hours of robotics-related data (VLA, VLN, Ego) on top of massive internet video, covering dexterous manipulation, robot locomotion, and first-person interaction. The model learns the relationship between actions and environmental change, not just surface texture.
Training: a multi-dimensional RL reward system aligns not only aesthetics, prompt following, and motion consistency, but also physical plausibility and task completion, using real-world video as preference signal.
Key benchmark: RBench (Peking University + ByteDance robot video benchmark) total score 0.620, above Wan 2.6 (0.607), Seedance 1.5 Pro (0.584), and NVIDIA Cosmos 3 Super (0.581). Ranked first on Physics-IQ Verified (physical phenomenon generation and prediction).
LingBot-World 2.0: 14B real-time interactive world model
A world model's core is simulating physical-world evolution. LingBot-World 2.0 supports:
720p/60fps is "game-grade" fidelity; one hour of stable output means world models have moved from short clips to long-horizon interactive simulation.
Why It Matters
1. A same-stack trio: engineering value beyond any single model
Individually: LingBot-VLA 2.0 is one open 6B VLA among many (π0.5, GR00T N1.7); LingBot-Video competes with Wan and Cosmos; LingBot-World 2.0 resembles GAIA-1 and UniSim. But released together, same-lineage and same-stack:
This same-stack design is what NVIDIA Cosmos aimed for but never fully delivered openly: Cosmos 1.0 was mostly a world model, Cosmos 2.0 added video generation, but VLAs still depended on external work (GR00T). This is the first complete, engineering-coherent open-sourcing of the full trio.
2. Cosmos对标, fully open
NVIDIA Cosmos positioned itself as embodied-AI infrastructure but has clear commercial boundaries: open base-model weights, but paid training data, pipelines, and post-training toolchains.
The LingBot trio is Apache-2.0 throughout — weights, code, and technical reports fully open for VLA, Video, and World. Any embodied-AI company, robotics firm, or research lab can commercially use this infrastructure for free — no NVIDIA payments, no NDAs. A huge lever for smaller companies, academia, and the open-source community.
For China's embodied-AI industry, this is close to a sovereignty milestone: previously the choice was NVIDIA Cosmos (dependence on a US ecosystem) or training from scratch (prohibitively costly). Now there is an Apache-2.0, world-class alternative.
3. An engineering exemplar of data scale and processing
60,000 hours plus 70,000 hours — the numbers are less surprising than the computability of the data:
LingBot-Video's data-profiling engine classifies internet video by physical consistency, keeping only what helps embodied AI. This is not a "big data" story but a demonstration of quantifiable, filterable, analyzable data — a standard worth adopting industry-wide.
4. "Paper + infrastructure" alongside Unitree's Nature paper
Same-week embodied-AI news:
Risks and Things to Watch
Honest caveats:
1. 6B + 130ms inference sounds great, but 130ms is 10 denoising steps on an RTX 4090D. At 25 steps, batching, or complex multi-step tasks, latency rises significantly. Real production end-to-end latency needs more measurement. 2. GM-100 is a high-quality, Cognition-style in-house evaluation, not a community-recognized SOTA benchmark. VLA lacks a SWE-bench equivalent. Leading on GM-100 doesn't guarantee leadership on BridgeData, RT-1, OpenVLA, etc. 3. The world model's "1-hour stable output" is a test condition, not a production condition. Real multi-user interaction, long-horizon tasks, and complex physical events will degrade stability — a shared weakness of all current world models. 4. Apache-2.0 + a Chinese team carries commercialization tension. The license permits closed-source enhanced forks without upstream contributions — a legal reality, not an accusation against Robbyant, but an open question for its commercial path.
Overall, the LingBot triple release is 2026's most infrastructure-significant embodied-AI open-source event: not "another open model" but a complete, same-lineage, same-stack technical stack open-sourced in one day — an organizational engineering capability itself worth tracking. For China's embodied-AI industry, it is the first proof that embodied-AI infrastructure need not be NVIDIA's alone.