China’s AlphaBrain Platform Launches First Full‑Stack Open‑Source Brain‑Like VLA

In April, Tesla disclosed patents for its Optimus humanoid robot, prompting a Chinese response: AI² Robotics released the AlphaBrain Platform, a one‑stop, open‑source embodied‑intelligence stack that spans data collection, model training, architecture design, and testing.

The platform offers a complete technical pipeline—including state‑of‑the‑art world models, brain‑like VLA models, plug‑and‑play composability, and a unified benchmark—previously confined to top‑tier labs, now freely available for developers worldwide.

NeuroVLA , the first open‑source brain‑like VLA model, introduces a spiking neural network (SNN) action head that mimics neuronal pulse firing. This enables online self‑adaptive learning during deployment without back‑propagation, while the GRU‑FiLM refinement module conditionally corrects SNN outputs to markedly improve motion precision.

To combat catastrophic forgetting, the platform provides a cross‑architecture continuous‑learning algorithm. By fine‑tuning only 6% of VLM backbone parameters with LoRA, memory consumption drops 60%. An experience‑replay buffer automatically replays old‑task samples, preserving prior skills. The approach has been verified on QwenGR00T and LLamaOFT, demonstrating true cross‑architecture compatibility.

The low‑cost generalization strategy employs a novel RL‑Token training architecture. A two‑stage process freezes the VLA backbone in the second stage, training a lightweight RL module only, which reduces compute to 3.5% of the original cost. Additionally, a 50% dropout of reference actions prevents actor degradation and encourages autonomous exploration.

The platform’s pluggable world‑model architecture natively integrates NVIDIA Cosmos Policy weights and supports seamless switching among Meta’s V‑JEPA, NVIDIA’s Cosmos Predict, and Alibaba’s Wan. All world models share a DiT action decoder and automatically adapt to each provider’s multimodal text encoder, allowing developers to compare model performance with minimal configuration changes.

AlphaBrain aligns with the latest embodied‑robotics benchmarks—LIBERO, LIBERO‑plus, RoboCasa, and RoboCasa365—offering a unified evaluation portal that automates the full inference lifecycle, supports WebSocket services, BF16 acceleration, remote deployment, and VLA+VLM joint training, thereby streamlining rigorous performance testing.

Beyond software, AI² Robotics supplies hardware such as the AlphaBot 2 robot, rated for 50,000 hours of fault‑free operation and produced at a scale of thousands per year. The company has secured large commercial orders (e.g., 1,000 units from a major panel manufacturer) and deployed modular embodied‑intelligence service spaces in retail, illustrating the platform’s end‑to‑end practicality.