On July 8-9, 2026, Robbyant (灵波科技), a subsidiary of Ant Group, open-sourced three embodied AI foundation models in a single coordinated release:
| Model | Params | Role | Key Benchmark | |-------|--------|------|---------------| | LingBot-VLA 2.0 | 6B | Vision-language-action model outputting robot actions directly | Leads π0.5 on GM-100 bimanual tasks | | LingBot-Video | 30B (3B activated) | Video foundation model for embodied AI | RBench 0.620, #1 on Physics-IQ Verified | | LingBot-World 2.0 | 14B | Real-time interactive world model | Stable 720p/60fps for 1 hour |
Together they form a "VLA + video foundation model + world model" stack — positioned against NVIDIA Cosmos, but fully open-source under Apache-2.0. The engineering significance lies less in any single model than in the three being released together, from one team, on one shared stack.
LingBot-VLA 2.0: A 6B Model Across 20 Robot Morphologies
VLA models map camera images and language instructions to robot actions. Prior VLA models suffered two problems: per-morphology training (different joint definitions) and slow inference (>200ms for most).
LingBot-VLA 2.0's solutions:
Unified action representation. A 55-dimensional vector unifies state and action across all morphologies:
- Arm joint positions: 14
- End-effector pose: 14 (XYZ + quaternion × 2)
- Gripper positions: 2
- Finger joint positions: 12
- Waist: 4
- Head: 2
- Mobile base signals: 3
- Reserved: 4
- Architecture: DiT + MoE, 30B total / 3B activated, ~3× more inference-efficient than dense at comparable scale.
- Data: A data profiling engine added ~70,000 hours of robot-relevant data (VLA, VLN, Ego), covering dexterous manipulation and first-person interaction — learning action-environment causality, not just texture.
- Training: Multi-dimensional RL rewards aligned on physical plausibility and task completion, with real-world video as preference signal, beyond aesthetics and prompt following.
- Benchmarks: RBench total 0.620 (above Wan 2.6 at 0.607, Seedance 1.5 Pro at 0.584, NVIDIA Cosmos 3 Super at 0.581); ranked first on Physics-IQ Verified.
- Rich character actions (casting, attacking, jumping) and text-driven events
- Built-in Pilot Agent + Director Agent for continuous world evolution
- Multi-user simultaneous interaction
- Stable 720p/60fps real-time output with no quality decay for up to 1 hour
- 130ms latency is measured on an RTX 4090D at 10 denoising steps; 25-step inference, batching, or multi-step tasks will raise latency. Production end-to-end timing needs more testing.
- GM-100 is not a community-standard benchmark. VLA lacks a SWE-bench equivalent; leadership on GM-100 does not guarantee superiority on BridgeData, RT-1, or OpenVLA settings.
- The 1-hour world model run is a test condition, not a production one. Stability under real multi-user interaction and complex physics will degrade — a weakness shared by all current world models.
- Apache-2.0 permits closed-source commercialization. Companies can build proprietary enhanced versions without reciprocating — a legal reality and open commercialization question, not a criticism of Robbyant itself.
Robots lacking a component get zeros in the corresponding dimensions. One model can therefore control single arms, bimanual setups, humanoids, and mobile robots.
MoE action experts. Sparse MoE replaces feed-forward layers — one shared expert plus routed experts with top-K activation per token. At equal activated parameters, training loss and validation action error beat dense baselines.
Data quality engineering. 60,000 hours of pretraining data: 50,000 hours of robot trajectories across 20 configurations, plus 10,000 hours of egocentric human video. The filtering pipeline explicitly measures third-order jerk and velocity/acceleration Z-scores, filtering static or jittering samples. Video is validated against robot states via URDF reconstruction; hand videos use MANO pose reconstruction.
Dual-query distillation. Two learnable queries attend to current (Q_t) and future (Q_{t+T}) observations, supervised by two teachers: LingBot-Depth (geometry) and DINO-Video (temporal dynamics).
Inference runs at 130ms on an RTX 4090D (10 denoising steps). On GM-100 (100 bimanual tasks), it scores 66.2/34.4 (progress/success) on AgileX Cobot Magic, beating GR00T N1.7, π0.5, and LingBot-VLA 1.0, and outperforms π0.5 on long-horizon mobile manipulation (Astribot S1 fridge sorting, Cobot Magic-ARX X5 stove cleaning), in-domain and OOD.
LingBot-Video: A 30B MoE Video Model Aimed at Robotics
LingBot-World 2.0: A 14B Real-Time Interactive World Model
The 1-hour stable horizon marks a shift from short video clips to long-horizon interactive simulation.
Why It Matters
1. One stack, three models. Individually, each competes with existing work (π0.5/GR00T, Wan/Cosmos, GAIA-1/UniSim). Together they unlock: shared data (robot trajectories teach the video model physical consistency; generated video augments VLA training), simulation environments for closed-loop VLA training via World 2.0, and shared backbones (Qwen3-VL-4B for VLA, shared ViT features, Pilot Agent reusing VLM instruction understanding). This is the goal NVIDIA Cosmos pursued but did not fully open-source — Cosmos 1.0 was a world model, 2.0 added video generation, but VLA remained external (GR00T).
2. Full Apache-2.0 openness vs. Cosmos. NVIDIA Cosmos opens base weights but keeps data, pipelines, and post-training tooling commercial. LingBot's three models ship weights, code, and technical reports fully open — free commercial use for embodied AI startups, robotics companies, and academia, without NVIDIA licensing. For China's embodied AI industry, it is the first demonstrated world-class open-source alternative to NVIDIA infrastructure.
3. Computable data as an engineering exemplar. The filtering pipeline measures jerk, velocity/acceleration Z-scores, URDF video consistency, and MANO hand reconstruction; LingBot-Video's profiling engine classifies internet video by physical consistency. This is a quantifiable, filterable, analyzable data paradigm worth industry-wide adoption.
4. A two-track moment. The same week brought a Nature paper on Unitree G1 (in-vivo surgery, July 8) alongside these three open-source releases — application-layer validation plus infrastructure-layer open-sourcing arriving together, signaling embodied AI's shift from paper-driven to engineering-driven iteration.