Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

NVIDIA Alpamayo

NVIDIA's open driving VLA models that plan a path and explain it, with sim, RL and data

License

Apache 2.0

Code: Apache-2.0, in the Developer Hub and in every model, simulator and tool repository; the dataset devkit is MIT. Model weights: OpenMDW-1.1 for all four checkpoints. It allows use, changes and redistribution, commercial use included, if the licence text and notices are kept, and it ends for anyone who sues claiming the model infringes a patent or copyright. Alpamayo 1 and 1.5: their model cards also say the models are ready for non-commercial use, with commercial licensing on request, which conflicts with OpenMDW-1.1; ask NVIDIA before commercial use. Dataset: the Physical AI Autonomous Vehicles dataset is under the NVIDIA Autonomous Vehicle Dataset License, gated on Hugging Face: internal development of autonomous vehicles and driver assistance using NVIDIA technology only, revocable, no redistribution or derived datasets, no identifying people, and it ends 12 months after download. Checked 2026-11-11 against the LICENSE files and READMEs, the Hugging Face model and dataset cards, and openmdw.ai.

Overview

NVIDIA Alpamayo is NVIDIA's open platform for research on reasoning-based autonomous driving. Its core is a family of vision-language-action models that read video from the cameras around a car, its recent motion and a navigation instruction, and return the path for the next 6.4 seconds together with a written explanation of the decision, which NVIDIA calls Chain-of-Causation reasoning. Four checkpoints are open: Alpamayo 1 and Alpamayo 1.5 at 10B parameters, Alpamayo 2 Super at 34B, and Alpamayo 2 Edge, a 2B-class model released on 10 November 2026 for running inside the vehicle on NVIDIA DRIVE AGX Thor. Around them sit AlpaSim for closed-loop driving simulation, AlpaGym for reinforcement learning in that simulator, a pipeline that writes reasoning labels for driving clips, the Physical AI Autonomous Vehicles dataset, and the Alpamayo Developer Hub with fine-tuning, reinforcement learning, distillation and quantization recipes. The code is Apache-2.0 and the model weights are OpenMDW-1.1; the dataset has its own, narrower NVIDIA licence.

NVIDIA announced Alpamayo at CES in January 2026 as a family of open models, simulation tools and datasets for autonomous vehicles, and has added to it through the year. The Alpamayo Developer Hub on GitHub is the family's home: it points to every model, simulator and dataset repository and holds the post-training recipes.

Every Alpamayo model is a VLA model built the same way. A VLM backbone from NVIDIA's Cosmos family reads several camera views over a short time window, the car's recent motion and a text instruction, and writes a short explanation of what it sees and what it will do: the Chain-of-Causation (CoC) trace, which links the decision to the objects and events that cause it. A diffusion-based action expert then turns that into a trajectory: 64 future poses at 0.1-second intervals, covering 6.4 seconds. The method is described in NVIDIA's Alpamayo-R1 paper. Four checkpoints are open:

- Alpamayo 1 Nano, first released as Alpamayo-R1 (weights): 10B parameters, an 8.2B Cosmos-Reason backbone and a 2.3B action expert, four front and side cameras, at least 24 GB of GPU memory. Its repository now receives only limited maintenance. - Alpamayo 1.5 Nano (weights): 10B parameters on Cosmos-Reason2, post-trained with reinforcement learning, and adds navigation instructions, a variable number of cameras and answers to questions about the scene. - Alpamayo 2 Super (weights): 34B parameters, a 32B Cosmos 3 backbone and a 2.3B action expert, with surround cameras, high-level decisions such as yield, lane change and stop, 2D grounding and automatic labelling. NVIDIA presents it as a teacher model for training smaller in-car models. - Alpamayo 2 Edge (weights), released on 10 November 2026: a 2.45B backbone from Cosmos 3 Edge, which is based on Nemotron, and a 0.2B action expert. It reads six cameras (22 frames), 20 past poses and a navigation instruction, writes its reasoning only when a situation calls for it, can answer free-form questions about the scene, and needs at least 16 GB of GPU memory for its BF16 examples, with a download of about 5.3 GB.

NVIDIA's Alpamayo 2 Edge blog post says the model was distilled from reasoning labels written by Alpamayo 2 Super and then trained with closed-loop reinforcement learning in AlpaGym, which improved its closed-loop driving by 48.9% over its best checkpoint without that step. After FP8 calibration, TensorRT engines and DFlash2 speculative decoding, NVIDIA reports 160 to 190 ms from camera pixels to trajectory on DRIVE AGX Thor, the computer NVIDIA builds for cars.

Around the models, AlpaSim is an open simulator for testing driving policies in closed loop, where each decision changes what the car sees next. It runs the driver, renderer, traffic, controller and physics as separate services that can sit on different GPUs, renders real recorded scenes through NuRec neural reconstruction or a generative video model, and also runs third-party policies. AlpaGym adds reinforcement learning on top of AlpaSim with Cosmos-RL and GRPO, using rewards for collisions, leaving the road and progress. The CoC auto-labeler finds the moments where a driver makes a decision in a clip and writes CoC traces for them with a VLM.

The Physical AI Autonomous Vehicles dataset holds 1,700 hours of driving from 25 countries and more than 2,500 cities, in 306,151 clips of 20 seconds, with seven cameras on every clip, lidar on 298,325 clips and radar on 160,760; a Python devkit reads it. Access requires accepting NVIDIA's licence on Hugging Face, and the model examples load their input from it.

The Developer Hub's recipes cover supervised fine-tuning of all four models, reinforcement learning with Cosmos-RL and GRPO for Alpamayo 1 and 1.5, distillation into smaller models, draft models for speculative decoding, and FP8 and NVFP4 quantization of Alpamayo 1.5. NVIDIA says the newest models were trained on about 115,000 hours of multi-camera driving video with about 3,700,000 CoC traces, and Alpamayo 1 and 1.5 on 80,000 hours.

Why it matters

Most open self-driving software is either a modular stack or, like Vision Pilot from the Autoware Foundation, Level 2 driver assistance from a single camera. Alpamayo is a different thing: end-to-end VLA driving models that go from camera video to a path in one network and say in words why they chose it, released with weights, inference code, training recipes, a closed-loop simulator, a reinforcement learning framework and a large dataset. That lets researchers study and adapt reasoning-based driving without first building the data, simulation and training around it. It is also one of the clearest public examples of a foundation model carried into a safety-critical vehicle task: it comes from the same NVIDIA Cosmos family as Cosmos Predict 2.5 and Cosmos Policy, and pairs a VLM with a diffusion action model as robot VLAs such as NVIDIA Isaac GR00T do. With Alpamayo 2 Edge the family reaches a size that runs on a single workstation GPU and, by NVIDIA's account, inside a car.

Problem it solves

End-to-end driving models that learn from data alone handle common traffic well but are brittle in rare situations and give no reason for what they do, which makes them hard to trust and to validate. Building an alternative needs large driving datasets, reasoning labels, a closed-loop simulator and a training setup, which few groups outside large companies have. Alpamayo releases all of these together: models that output both a trajectory and a written cause for it, the tools to fine-tune and test them in closed loop, and a smaller model meant to run within an in-car compute and latency budget.

Key takeaways

  • Four open driving VLA checkpoints: Alpamayo 1 and 1.5 at 10B, Alpamayo 2 Super at 34B and Alpamayo 2 Edge at 2B class
  • Each outputs 64 future poses over 6.4 seconds together with a written Chain-of-Causation explanation
  • Alpamayo 2 Edge, released 10 November 2026, runs its examples in 16 GB of GPU memory and targets NVIDIA DRIVE AGX Thor
  • AlpaSim and AlpaGym give closed-loop simulation and reinforcement learning for driving policies
  • The Developer Hub has recipes for fine-tuning, reinforcement learning, distillation, speculative decoding and quantization
  • Code Apache-2.0, weights OpenMDW-1.1; the driving dataset has a narrower NVIDIA licence

Who it is for

  • Autonomous-driving researchers working on end-to-end and reasoning-based planning
  • AV teams who want a teacher model to distil, or a compact model to fine-tune on their own fleet data
  • Simulation and reinforcement learning researchers who need a closed-loop driving environment
  • Researchers studying vision-language-action models and explainable decisions in safety-critical settings

Limitations

  • Built for cars: the models expect a vehicle's multi-camera rig, ego-motion history and calibrated cameras, and nothing upstream shows them used on other robots
  • Research tools, not a driving stack: NVIDIA says Alpamayo 1 and 1.5 are for research and evaluation and have not been validated to automotive grade
  • The Alpamayo 1 and 1.5 model cards list the weights as OpenMDW-1.1 but also say they are ready for non-commercial use with commercial licensing on request; ask NVIDIA before commercial use
  • The driving dataset is gated and licensed only for internal autonomous-vehicle development using NVIDIA technology, without redistribution, and the licence ends 12 months after download
  • GPU memory: at least 16 GB for Alpamayo 2 Edge, 24 GB for Alpamayo 1 and 1.5; Alpamayo 2 Super was tested on one 80 GB H100, and real-time in-car use needs DRIVE AGX Thor
  • The 2 Edge deployment recipe for DRIVE AGX Thor is described in NVIDIA's blog, but the hub README of 11 November 2026 still lists it as planned
  • Benchmark, simulation and in-car results are NVIDIA's own evaluations and partner statements
  • Alpamayo 1 now gets only limited maintenance, and AlpaGym is in early development and supports only Alpamayo 1.5

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.