[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-1737":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":17,"stars7d":18,"stars30d":19,"stars90d":16,"forks30d":16,"starsTrendScore":20,"compositeScore":21,"rankGlobal":10,"rankLanguage":10,"license":22,"archived":23,"fork":23,"defaultBranch":24,"hasWiki":25,"hasPages":25,"topics":26,"createdAt":10,"pushedAt":10,"updatedAt":27,"readmeContent":28,"aiSummary":29,"trendingCount":16,"starSnapshotCount":16,"syncStatus":30,"lastSyncTime":31,"discoverSource":32},1737,"AlphaBrain","AlphaBrainGroup\u002FAlphaBrain","AlphaBrainGroup","The Comprehensive Toolkit for Embodied AI Models","",null,"Python",202,32,168,1,0,5,11,26,15,62.66,"Other",false,"main",true,[],"2026-06-12 04:00:11","\u003Cdiv align=\"center\">\n\n# AlphaBrain\n\n### A Modular Open-Source Framework for Embodied Intelligence Research\n\n[![License: MIT](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-MIT-yellow.svg)](https:\u002F\u002Fopensource.org\u002Flicenses\u002FMIT)\n[![Docs](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FDocs-Online-green.svg)](https:\u002F\u002Falphabraingroup.github.io\u002FAlphaBrain\u002F)\n[![HuggingFace](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97-Models-orange.svg)](https:\u002F\u002Fhuggingface.co\u002FAlphaBrainGroup)\n[![WeChat](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FWeChat-Group-07C160.svg)](assets\u002Fwechat.jpg)\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"assets\u002Fmain_fig.png\" width=\"100%\" alt=\"AlphaBrain Architecture Overview\"\u002F>\n\u003C\u002Fp>\n\n**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.\n\n[Quick Start & Documentation](#-quick-start--documentation) · [Key Features](#-key-features) · [Community](#-community) · [Citation](#-citation)\n\n\u003C\u002Fdiv>\n\n---\n\n## Highlights\n\n\u003Ctable>\n\u003Ctr>\n\u003Ctd width=\"50\" align=\"center\">🧠\u003C\u002Ftd>\n\u003Ctd>\u003Cb>Brain-Inspired VLA (NeuroVLA)\u003C\u002Fb> — The first open-source biologically-inspired VLA model, achieving \u003Cb>SOTA on brain-inspired control tasks\u003C\u002Fb>. Integrates spiking neural networks (SNN) with STDP learning rules, advancing embodied intelligence toward biological brain learning mechanisms.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd width=\"50\" align=\"center\">🔄\u003C\u002Ftd>\n\u003Ctd>\u003Cb>Cross-Architecture Continual Learning\u003C\u002Fb> — 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.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd width=\"50\" align=\"center\">🎯\u003C\u002Ftd>\n\u003Ctd>\u003Cb>RLT_a Training Paradigm\u003C\u002Fb> — The first open-source VLA training architecture based on \u003Cb>RL Token\u003C\u002Fb>, a novel architecture that compresses VLA hidden states through an information bottleneck, followed by off-policy Actor-Critic reinforcement learning.\u003C\u002Ftd>\n\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd width=\"50\" align=\"center\">🌍\u003C\u002Ftd>\n\u003Ctd>\u003Cb>Native World Model Integration\u003C\u002Fb> — The first open-source VLA to natively integrate \u003Cb>Cosmos Policy\u003C\u002Fb> original weights, supporting flexible world model switching across Cosmos 2 \u002F 2.5, Wan 2.2, and V-JEPA 2.1.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd width=\"50\" align=\"center\">📊\u003C\u002Ftd>\n\u003Ctd>\u003Cb>Comprehensive Benchmark Suite\u003C\u002Fb> — Full adaptation to the latest embodied benchmarks with open-source support for \u003Cb>long-horizon task execution and memory\u003C\u002Fb>: LIBERO, LIBERO-plus, RoboCasa, RoboCasa365 and more to come.\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\n---\n\n## 🚀 Quick Start & Documentation\n\nFull setup, training, evaluation, and deployment instructions live in our documentation site. Step-by-step guides, configuration references, and troubleshooting notes are all maintained there.\n\n👉 **[AlphaBrain Documentation →](https:\u002F\u002Falphabraingroup.github.io\u002FAlphaBrain\u002F)**\n\n---\n\n## 🔬 Key Features\n\nAlphaBrain delivers five core capabilities on a single stack: the **VLA framework family** as the base, with **NeuroVLA \u002F RLT_a \u002F Continual Learning \u002F World Model** as composable capability modules. All capabilities share the same trainer, config system, and inference interface.\n\n### VLA Frameworks\n\n| Framework | Action Decoding | Typical Use |\n|:----------|:----------------|:------------|\n| **OFT** | MLP action head, parallel continuous decoding | Fast prototyping, baseline alignment |\n| **GR00T** | System1 + Flow-Matching DiT System2 | High-precision manipulation, long-horizon planning |\n| **PI** | Flow-Matching action prediction | Diffusion-style policies |\n| **LoRA Adapter** | Rank-32 adapters on all-linear layers, applied to all supported frameworks | Parameter-efficient fine-tuning |\n| **NeuroVLA** | Bio-inspired spiking + STDP | Brain-inspired control |\n| **CosmosPolicy** | Latent-space video diffusion | World-model-native policy |\n\n### Brain-Inspired VLA (NeuroVLA + STDP)\n\nNeuroVLA integrates spiking neural networks with biological learning rules into the VLA pipeline:\n\n- **QFormer** extracts layer-wise action-relevant features from VLM hidden states;\n- **SNN Action Head** with Leaky Integrate-and-Fire (LIF) neurons for spike-based action prediction;\n- **R-STDP Training** — Reward-Modulated Spike-Timing-Dependent Plasticity, supporting both hybrid (backprop + STDP) and pure STDP modes;\n- **Online STDP** — Test-time adaptation with zero backpropagation, using self-supervised reward signals from environment interaction.\n\n### RLT_a Online RL Fine-tuning\n\nA novel architecture that compresses VLA hidden states through an information bottleneck, followed by off-policy Actor-Critic reinforcement learning:\n- **Encoder-Decoder**: Extracts a compact action token from the VLA's internal features to serve as the state representation for RL.\n- **Two-Phase Training**: An initial adaptation stage to expose the action token → RL fine-tuning with a frozen VLA.\n- **Low Resource Requirements**: The actual reinforcement learning gradient update phase involves a highly lightweight parameters.\n\n### Continual Learning\n\nSequential task acquisition with a pluggable family of CL algorithms,\nselectable from YAML — no code edits needed to switch methods:\n\n- **Algorithms** — `ER` (Experience Replay) and `MIR` (Maximally Interfered\n  Retrieval, NeurIPS 2019) ship in-tree; both inherit a common\n  `CLAlgorithm` hook interface (`observe`, `modify_batch`,\n  `compute_penalty`, `after_backward`, `on_task_start`, `on_task_end`).\n- **Benchmarks** — `LIBERO-{Spatial,Object,Goal,Long}` ship as\n  first-class CL streams; arbitrary LeRobot-format mixtures supported\n  via YAML.\n- **Metrics** — one-command aggregator emits **ASR \u002F BWT (Lopez-Paz &\n  Ranzato 2017) \u002F F (Chaudhry et al. 2018)** from a T×T evaluation\n  matrix: `python scripts\u002Frun_continual_learning_scripts\u002Fcompute_cl_matrix_metrics.py \u003Cmatrix_dir>`.\n- **LoRA integration** — parameter-efficient fine-tuning (~6% trainable\n  params, ~3× memory savings); full-parameter variants also supported.\n- **Cross-architecture adaptation** — the same CL algorithm drops\n  directly onto QwenGR00T, NeuroVLA, LlamaOFT, PaliGemmaOFT, etc.\n\nDetailed quickstart, leaderboard, CLI flags, and the 77 % MIR-on-LIBERO-Goal\nrecipe: [`scripts\u002Frun_continual_learning_scripts\u002FREADME.md`](scripts\u002Frun_continual_learning_scripts\u002FREADME.md).\n\n### World Model Integration\n\nNative support for 4 world model backbones plus full CosmosPolicy finetuning:\n\n| Backbone | Params | Mode Name | Text Encoder |\n|:---------|:-------|:----------|:-------------|\n| V-JEPA 2.1 | ~1.8B | `world_model_vjepa` | T5-small |\n| Cosmos Predict 2.5 | ~2.1B | `world_model_cosmos` | Reason1-7B |\n| Cosmos Predict 2 | ~2.1B | `world_model_cosmos2` | T5-XXL |\n| Wan 2.2 | ~5B | `world_model_wan` | UMT5-XXL |\n\n---\n\n### Benchmarks\n\n| Benchmark | Tasks | Highlights | Path |\n|:----------|:------|:-----------|:-----|\n| **LIBERO** | Spatial \u002F Object \u002F Goal \u002F Long-horizon | Core evaluation suite, 4 task suites | `benchmarks\u002FLIBERO\u002F` |\n| **LIBERO-plus** | Robustness (Camera, Robot, Language, Light, etc.) | Zero-shot generalization testing | `benchmarks\u002FLIBERO-plus\u002F` |\n| **RoboCasa** | Tabletop & kitchen manipulation | Real-world scene diversity | `benchmarks\u002FRobocasa_tabletop\u002F` |\n| **RoboCasa365** | 365-day kitchen task collection | Large-scale daily tasks | `benchmarks\u002FRobocasa365\u002F` |\n| ... | | |\n\n---\n\n## 🤝 Community\n\nWe 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.\n\n| Channel | Link |\n|:--------|:-----|\n| GitHub Issues | [Report bugs & request features](https:\u002F\u002Fgithub.com\u002FAlphaBrainGroup\u002FAlphaBrain\u002Fissues) |\n| HuggingFace | [Models](https:\u002F\u002Fhuggingface.co\u002FAlphaBrainGroup) |\n| WeChat Group | [Scan the QR code to join](assets\u002Fwechat.jpg) |\n\n### Acknowledgments\n\nAlphaBrain is mainly forked from [starVLA](https:\u002F\u002Fgithub.com\u002FstarVLA\u002FstarVLA) 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:\n\n- [starVLA\u002FstarVLA](https:\u002F\u002Fgithub.com\u002FstarVLA\u002FstarVLA)\n- [openvla\u002Fopenvla](https:\u002F\u002Fgithub.com\u002Fopenvla\u002Fopenvla)\n- [moojink\u002Fopenvla-oft](https:\u002F\u002Fgithub.com\u002Fmoojink\u002Fopenvla-oft)\n- [Physical-Intelligence\u002Fopenpi](https:\u002F\u002Fgithub.com\u002FPhysical-Intelligence\u002Fopenpi)\n- [NVIDIA\u002FIsaac-GR00T](https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FIsaac-GR00T)\n- [QwenLM\u002FQwen3-VL](https:\u002F\u002Fgithub.com\u002FQwenLM\u002FQwen3-VL)\n- [nvidia-cosmos\u002Fcosmos-predict2.5](https:\u002F\u002Fgithub.com\u002Fnvidia-cosmos\u002Fcosmos-predict2.5)\n- [Wan-Video\u002FWan2.2](https:\u002F\u002Fgithub.com\u002FWan-Video\u002FWan2.2)\n- [Lifelong-Robot-Learning\u002FLIBERO](https:\u002F\u002Fgithub.com\u002FLifelong-Robot-Learning\u002FLIBERO)\n- [robocasa\u002Frobocasa](https:\u002F\u002Fgithub.com\u002Frobocasa\u002Frobocasa)\n- [guoweiyu\u002FNeuroVLA](https:\u002F\u002Fgithub.com\u002Fguoweiyu\u002FNeuroVLA)\n\n\n---\n\n## 📝 Citation\n\n```bibtex\n@software{AlphaBrain2026,\n  title     = {AlphaBrain: a Modular Open-Source Framework for Embodied Intelligence Research},\n  author    = {AlphaBrain Community},\n  year      = {2026},\n  url       = {https:\u002F\u002Fgithub.com\u002FAlphaBrainGroup\u002FAlphaBrain},\n  license   = {MIT},\n  doi       = {}\n}\n```\n\n---\n\n## 📄 License\n\nThis project is licensed under the [MIT License](LICENSE).\n\n\u003Cdiv align=\"center\">\n\u003Csub>Built with passion by the AlphaBrain Community upon starVLA\u003C\u002Fsub>\n\u003C\u002Fdiv>\n","AlphaBrain 是一个面向具身智能研究的模块化开源框架。它集成了多种VLA架构、世界模型骨干、生物启发的学习算法和强化学习范式，提供了一个可扩展的统一平台。该项目的核心功能包括首个开源的生物启发VLA模型（NeuroVLA），跨架构连续学习算法，基于RL Token的训练范式，以及原生集成Cosmos Policy的世界模型支持。此外，AlphaBrain还提供了全面的基准测试套件，适用于长时任务执行与记忆等场景。此工具包特别适合于需要探索或开发具身AI解决方案的研究人员和技术开发者使用。",2,"2026-06-11 02:45:44","CREATED_QUERY"]