Robbyant has released LingBot-VA 2.0, a foundation model for robot manipulation described as the first pretraining system built from the ground up around physical causality — not adapted from language or vision models. The release arrives the week before World AI Conference 2026 (WAIC) in Shanghai. On the RoboTwin 2.0 dual-arm benchmark, the model reaches 93.6% average success rate — a new state-of-the-art for this class.
Key takeaways
- 93.6% average on RoboTwin 2.0 (Clean: 93.8%, Randomized: 93.4%) — above LingBot-VA 1.0 (92.2%) and Motus (87.9%)
- New VAE tokenizer aligns visual semantics and latent actions in a single representation space, without any human labels
- Multi-Chunk Prediction (MCP): +29.7 pp success rate at 5,000 training steps, 2.3× convergence acceleration
- Foresight Reasoning: async pipeline raises control frequency from 35 Hz to 225 Hz, inference latency drops from 927 ms to 142 ms
- In-context learning from a single human demonstration — task transfer to a new object without fine-tuning
Why "natively embodied" is more than a tagline
Most VLA (Vision-Language-Action) systems originate from language or vision models trained on static text and image data. LingBot-VA 2.0 claims that from the very first training step, the model focuses on causal relations in the physical world — not on token prediction migrated to robot action space.
The central innovation is a redesigned VAE tokenizer. Rather than treating the visual channel and the action channel separately, the model introduces a single ViT autoencoder that aligns visual fragments (scene semantics) and latent actions (what the robot should execute) within the same latent geometry. Two auxiliary dynamics models — an IDM (Inverse Dynamics Model) extracting actions from adjacent frame pairs and an FDM (Forward Dynamics Model) predicting forward states — operate entirely without human annotations. The architecture learns directly from raw video.
The practical effect is that instruction comprehension and motion generation share the same representational space, reducing the gap between "understanding a command" and "producing a sequence of movements."
Multi-chunk prediction instead of next-frame supervision
Standard video-action training with teacher forcing predicts one chunk of frames from the immediately preceding chunk — the model mostly learns from short-range, visually similar pairs. LingBot-VA 2.0 introduces Multi-Chunk Prediction (MCP): three lightweight modules attached to the main DiT backbone, each responsible for predicting a progressively more distant video segment (three future chunks by default). The cascade forces latent variables to encode trajectory dynamics over a longer horizon.
RoboTwin results at 50 fps are clear: at 5,000 training steps, success rate improves by 29.7 percentage points over the baseline, and matching the quality of a full 45,000-step baseline requires only 20,000 steps — a 2.3× effective acceleration. At deployment, MCP has zero cost: the auxiliary modules are detachable after training.
Asynchronous thinking and acting: from 35 Hz to 225 Hz
The conventional control loop is sequential: observe → infer → execute. Any inference delay accumulates as actuation latency. LingBot-VA 2.0 decouples the two streams via Foresight Reasoning. The prediction stream continuously generates the next action from a projected world state. The execution stream runs the current action. When a real sensor observation returns, the model overwrites its prior forecasts in the KV-cache and corrects the trajectory — without pausing the control loop.
The inference optimization stack has four layers:
- consistency distillation — 927 ms → 466 ms
- FP8 compilation with TensorRT — 369 ms
- long-horizon attention via FlashInfer — 272 ms
- runtime overhead reduction — 142 ms
End-to-end acceleration reaches 6.5×, and the async control frequency rises from 35 Hz to 225 Hz.
MoE architecture and real-world results
Model capacity comes from a sparse MoE architecture borrowed from LingBot-Video, Robbyant's own video generation model. Visual experts run in sparse MoE mode on each DiT block; action experts remain denser. Routing follows an expansive scheme inspired by DeepSeek; load balancing is achieved without an auxiliary loss — via per-expert token-level micro-tuning instead of a global gradient multiplier.
In deployment demonstrations, the model controls a robot arm playing air hockey in real time, grasps potato chips without crushing them at precision levels requiring sub-threshold grip forces, coordinates two arms for assembly and tidying, and — after watching a single human demo — transfers the task schema to a new object without any fine-tuning.
Why this matters
LingBot-VA 2.0 fits a consensus now visible across multiple independent labs in 2026: robot control models should understand physical world dynamics, not merely map language tokens to servo signals. Cosmos from Google DeepMind, the Gemini Robotics series, and Physical Intelligence's work converge on the same conclusion from different starting points.
What differentiates LingBot-VA 2.0 is the combination of complementary mechanisms — latent alignment via VAE, multi-horizon prediction, and asynchronous inference — each with a measurable benchmark contribution. Particularly significant is the minimal gap between clean (93.8%) and randomized (93.4%) conditions: the model does not degrade under visual and physical perturbations, which is a serious challenge for imitation-based methods trained under controlled conditions.
What's next
- Robbyant plans a live demonstration at World AI Conference 2026 in Shanghai (July 2026); the full technical paper is available in the project's GitHub repository
- The next maturity test is deployment outside controlled lab environments — the company has indicated plans for internal production-platform trials with multiple tasks on a single deployed model
- Physical Intelligence, Google Gemini Robotics, and other VLA teams will sustain pressure for further benchmark improvements through the second half of 2026
Sources
- 机器之心 (Jiqizhixin) — 今天，具身智能大模型迈入「具身原生」的新阶段！
- Robbyant Technology — LingBot-VA 2.0 Project Page
- GitHub / Robbyant — Native Video-Action Pretraining for Generalizable Robot Control
