AlphaBrain

A Modular Open-Source Framework for Embodied Intelligence Research

AlphaBrain is an all-in-one, open-source community for embodied intelligence, built to be ready out of the box. We unifies multiple VLA architectures, world model backbones, biologically-inspired learning algorithms, and reinforcement learning paradigms under a single, extensible framework. AlphaBrain brings embodied AI within everyone’s reach.

Quick Start & Documentation · Key Features · Community · Citation

---

Highlights

🧠	Brain-Inspired VLA (NeuroVLA) — The first open-source biologically-inspired VLA model, achieving SOTA on brain-inspired control tasks. Integrates spiking neural networks (SNN) with STDP learning rules, advancing embodied intelligence toward biological brain learning mechanisms.
🔄	Cross-Architecture Continual Learning — The first open-source continual learning algorithm designed for cross-architecture VLA, breaking architecture compatibility bottlenecks and supporting universal adaptation and knowledge accumulation across different VLA models.
🎯	RLActionToken Training Paradigm — The first open-source VLA training architecture based on RL Token, a novel architecture that compresses VLA hidden states through an information bottleneck, followed by off-policy Actor-Critic reinforcement learning.
🌍	Native World Model Integration — The first open-source VLA to natively integrate Cosmos Policy original weights, supporting flexible world model switching across Cosmos 2 / 2.5, Wan 2.2, and V-JEPA 2.1.
📊	Comprehensive Benchmark Suite — Full adaptation to the latest embodied benchmarks with open-source support for long-horizon task execution and memory: LIBERO, LIBERO-plus, RoboCasa, RoboCasa365 and more to come.

---

🚀 Quick Start & Documentation

Full setup, training, evaluation, and deployment instructions live in our documentation site. Step-by-step guides, configuration references, and troubleshooting notes are all maintained there.

👉 AlphaBrain Documentation →

---

🔬 Key Features

AlphaBrain delivers five core capabilities on a single stack: the VLA framework family as the base, with NeuroVLA / RLActionToken / Continual Learning / World Model as composable capability modules. All capabilities share the same trainer, config system, and inference interface.

VLA Frameworks

Framework	Action Decoding	Typical Use
OFT	MLP action head, parallel continuous decoding	Fast prototyping, baseline alignment
GR00T	System1 + Flow-Matching DiT System2	High-precision manipulation, long-horizon planning
PI	Flow-Matching action prediction	Diffusion-style policies
Adapter	Lightweight Adapter decoding	Parameter-efficient fine-tuning
NeuroVLA	Bio-inspired spiking + STDP	Brain-inspired control
CosmosPolicy	Latent-space video diffusion	World-model-native policy

Brain-Inspired VLA (NeuroVLA + STDP)

NeuroVLA integrates spiking neural networks with biological learning rules into the VLA pipeline:

• QFormer extracts layer-wise action-relevant features from VLM hidden states;
• SNN Action Head with Leaky Integrate-and-Fire (LIF) neurons for spike-based action prediction;
• R-STDP Training — Reward-Modulated Spike-Timing-Dependent Plasticity, supporting both hybrid (backprop + STDP) and pure STDP modes;
• Online STDP — Test-time adaptation with zero backpropagation, using self-supervised reward signals from environment interaction.

RLActionToken Online RL Fine-tuning

A novel architecture that compresses VLA hidden states through an information bottleneck, followed by off-policy Actor-Critic reinforcement learning:

• Encoder-Decoder: Extracts a compact action token from the VLA's internal features to serve as the state representation for RL.
• Two-Phase Training: An initial adaptation stage to expose the action token → RL fine-tuning with a frozen VLA.
• Low Resource Requirements: The actual reinforcement learning gradient update phase involves a highly lightweight parameters.

Continual Learning

Experience-replay-based continual learning for sequential task acquisition:

• Incremental design — all changes are additive, no modification to base training code;
• LoRA integration — parameter-efficient fine-tuning (~6% trainable params, ~3× memory savings);
• Replay buffer with configurable per-task capacity;
• Cross-architecture adaptation — the same CL algorithm drops directly onto different VLA frameworks.

World Model Integration

Native support for 4 world model backbones plus full CosmosPolicy finetuning:

Backbone	Params	Mode Name	Text Encoder
V-JEPA 2.1	~1.8B	world_model_vjepa	T5-small
Cosmos Predict 2.5	~2.1B	world_model_cosmos	Reason1-7B
Cosmos Predict 2	~2.1B	world_model_cosmos2	T5-XXL
Wan 2.2	~5B	world_model_wan	UMT5-XXL

---

Benchmarks

Benchmark	Tasks	Highlights	Path
LIBERO	Spatial / Object / Goal / Long-horizon	Core evaluation suite, 4 task suites	benchmarks/LIBERO/
LIBERO-plus	Robustness (Camera, Robot, Language, Light, etc.)	Zero-shot generalization testing	benchmarks/LIBERO-plus/
RoboCasa	Tabletop & kitchen manipulation	Real-world scene diversity	benchmarks/Robocasa_tabletop/
RoboCasa365	365-day kitchen task collection	Large-scale daily tasks	benchmarks/Robocasa365/
...

---

🤝 Community

We welcome contributions from the community — including new frameworks, benchmark adapters, bug fixes, and improvements that achieve stronger benchmark performance. Outstanding contributors may be invited to join the community as core members. Every contribution matters.

Channel	Link
GitHub Issues	Report bugs & request features
HuggingFace	Models
WeChat Group	Scan the QR code to join

Acknowledgments

AlphaBrain is mainly forked from starVLA and stands on the shoulders of an incredible open-source ecosystem. We are deeply grateful to the authors and maintainers of the following projects, whose code, models, datasets, and ideas directly enabled this work:

• starVLA/starVLA
• openvla/openvla
• moojink/openvla-oft
• Physical-Intelligence/openpi
• NVIDIA/Isaac-GR00T
• QwenLM/Qwen3-VL
• nvidia-cosmos/cosmos-predict2.5
• Wan-Video/Wan2.2
• Lifelong-Robot-Learning/LIBERO
• robocasa/robocasa
• guoweiyu/NeuroVLA

---

📝 Citation

@software{AlphaBrain2026,
  title     = {AlphaBrain: a Modular Open-Source Framework for Embodied Intelligence Research},
  author    = {AlphaBrain Community},
  year      = {2026},
  url       = {https://github.com/AlphaBrainGroup/AlphaBrain},
  license   = {MIT},
  doi       = {}
}

---

📄 License

This project is licensed under the MIT License.

_{Built with passion by the AlphaBrain Community upon starVLA}