TurboVLA

τ₀-VLA is an open-source hierarchical robot foundation model and reference implementation for long-horizon manipulation. A memory-augmented high-level policy generates the next subtask and uses world-model-guided test-time computation to search over alternatives when additional reasoning is needed; a generalist low-level policy then executes the selected subtask across robot embodiments. The low-level policy combines a Qwen3.5 vision-language backbone with a Mixture-of-Transformers action expert trained through conditional flow matching, using a unified 40-dimensional state/action space and multimodal co-training on tens of thousands of hours of heterogeneous real-world robot data. The repository provides LeRobot-format data loading, prompting, masking and normalization, embodiment-specific adapters, training and post-training recipes, deployment contracts, a policy server, and open-loop evaluation tools. It includes an AgiBot World example subset and templates for other datasets and robots. Public v1 serving supports joint-control checkpoints; native end-effector data can be used for training but end-effector serving is not supported in that release. The reference environment uses Python 3.11, CUDA 12.8, and PyTorch 2.7.1. Code and model weights are released under the Apache License 2.0.

vLLM is an open-source library and inference-serving engine for deploying large language models on GPUs and other hardware accelerators. Originally developed in UC Berkeley's Sky Computing Lab, it is maintained by a broad community and provides an OpenAI-compatible API server, as well as Anthropic Messages API and gRPC support. Its serving stack manages attention key-value memory with PagedAttention, continuous batching, chunked prefill, prefix caching, streaming generation, and disaggregated prefill, decode, and encode. It also supports speculative decoding, quantization, optimized attention and GEMM/MoE kernels, automatic kernel generation and graph-level transformations, structured outputs, tool calling, multiple decoding algorithms, and distributed inference through tensor, pipeline, data, expert, and context parallelism. The project integrates with Hugging Face models and supports decoder-only, mixture-of-experts, hybrid state-space, multimodal, embedding, retrieval, reward, and classification models. It can run across NVIDIA, AMD, Intel, and other supported accelerators, as well as x86, ARM, and PowerPC CPUs, and can be installed with uv or pip or built from source.

mini-AGI is an experimental continual-learning, byte-level language model that trains from scratch on a CUDA-capable GPU with at least 8 GB of VRAM. It reads 256 byte values directly rather than using a conventional tokenizer, processes data one chunk at a time, and uses the same forward path for training and generation. Its architecture combines dense prelude blocks with a recurrent block applied up to 24 times, adaptive PonderNet-style halting, and per-application top-8 routing through a dynamically growing and pruning expert pool. Expert weights and Adam moments are stored as files on disk; a working set is paged through RAM and VRAM as needed, allowing the pool size to exceed available VRAM. The project includes Python commands for building corpora, reading files, streaming training, serving a local web interface, and inspecting or replicating continual-learning experiments. The repository describes the current model as toy-level rather than frontier-capable, and its weights are generated in the local weights directory rather than distributed with the source.

LightNav-0 is an open-source generalist embodied-navigation model from the Light Origins Team. It uses a pretrained Qwen3-VL backbone with a single egocentric RGB stream and natural-language instructions to control humanoid, quadruped, wheeled, and aerial robots, transferring across tasks, embodiments, and scenes without per-task or per-benchmark fine-tuning. The model extends the vocabulary with dual-channel pointing tokens and residual vector-quantized action tokens rather than adding navigation-specific prediction heads. At each step, it emits an affordance point representing a feasible local direction or waypoint and an object point representing the goal, followed by three action tokens that decode to ten future SE(2) waypoints for an embodiment-specific low-level controller. Its temporally aware history compressor samples older frames less frequently, pools them more coarsely, and preserves ordering with timestamp tokens while bounding the visual context. The repository includes a released checkpoint and action decoder, command-line prediction, a vLLM-based server with WebSocket streaming, a MuJoCo simulation demo, evaluation harnesses, and a ROS 2 deployment stack with adapters for the Unitree Go2 and LimX TRON 1. The project is released under the Apache License 2.0; the included EVT-Bench is separately licensed CC BY-NC-SA 4.0.

LingBot-Map is a feed-forward 3D foundation model from the Robbyant Team for streaming 3D reconstruction from image sequences and video. Its Geometric Context Transformer unifies coordinate grounding, dense geometric cues, and long-range drift correction through anchor context, a pose-reference window, and trajectory memory. The system uses paged KV-cache attention for streaming inference and supports interactive browser-based reconstruction with point-cloud visualization, keyframe intervals, windowed processing for long sequences, and sky masking. It also provides an offline rendering pipeline for image folders or video, along with evaluation pipelines for datasets including KITTI and Oxford Spires. The repository reports approximately 20 FPS inference at 518×378 resolution on sequences exceeding 10,000 frames, and documents PyTorch installation with FlashInfer as the recommended attention backend and native SDPA as a fallback.