Alibaba opens ABot models

Alibaba’s Amap unit has opened part of its ABot robotics stack, releasing the manipulation model ABot-M0 with code, weights and data after publishing the technical report in February. (github.com) (arxiv.org) Embodied artificial intelligence is software for robots that has to see, decide and act in the physical world, not just answer questions on a screen. Amap’s ABot line splits that job into manipulation, which is moving objects with robot arms, and navigation, which is moving through spaces toward goals. (arxiv.org 1) (arxiv.org 2) Amap says ABot-M0 was trained on more than 6 million trajectories and about 9,500 hours of interaction data drawn from six public datasets covering more than 20 robot morphologies. The project page says the model uses one architecture for single-arm and dual-arm tasks. (amap-cvlab.github.io) (arxiv.org) The model’s main technical claim is “Action Manifold Learning,” a method that predicts robot actions directly instead of first modeling noise and then removing it. In the paper, the authors say that shift improved decoding speed and policy stability in their experiments. (amap-cvlab.github.io) (arxiv.org) On navigation, Amap’s ABot-N0 project page says the model was trained with 16.9 million expert trajectories and 5.0 million reasoning samples across 7,802 simulated 3D scenes. The same page says ABot-N0 is designed to handle five tasks with one system, including point-goal, object-goal and instruction-following navigation. (amap-cvlab.github.io) (arxiv.org) Amap’s April 15 announcement tied those models to a product push, saying its first embodied robot, a quadruped, is set to debut at the 2026 Beijing E-Town Half Marathon and Humanoid Robot Half Marathon. The company also said quadruped and humanoid robots are on the division’s roadmap. (prnewswire.com) The company’s pitch is that open weights and training pipelines can lower the cost of entry for labs that do not have their own large robot datasets. The GitHub repository says ABot-M0 training code, pre-trained weights and data became available on March 27, 2026, after inference code and weights were posted on February 27. (github.com) Amap also says ABot-M0 posted state-of-the-art results on Libero, Libero-Plus and RoboCasa, while ABot-N0 led seven navigation benchmarks including CityWalker and HM3D-OVON. Those are company-reported benchmark results from project pages, papers and an April 15 press release, and outside replication was not cited in the materials reviewed. (prnewswire.com) (amap-cvlab.github.io 1) (amap-cvlab.github.io 2) Alibaba’s move lands as robotics groups race to build “one brain, many forms” systems that can transfer skills across different machines. Amap’s papers describe that as the central problem: too many robot datasets, too many incompatible action formats, and too many models built for one task at a time. (arxiv.org 1) (arxiv.org 2) For now, the clearest change is practical: ABot-M0 is no longer just a paper claim, but a downloadable base model with code and weights that other teams can test on their own hardware. (github.com)