Hugging Face AI Updates: November 7, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Ai2’s ToyBlocks Opens a Simulated Tabletop for Robot Learning, Where Sim-Trained Skills Hit 81.1% on a Real Franka Arm
Ai2 released ToyBlocks on the Hugging Face blog, the first of what it calls FoundationWorlds: controlled simulated environments paired with automatic task and demonstration generators and behavior metrics, so researchers can vary training data or model design and trace why a robot fails. ToyBlocks is a robot arm with lettered and pictured blocks on a table, with domain randomization of block size, table height and other physical properties; short-horizon skills such as pushing and flipping are learned with reinforcement learning and motion planning in NVIDIA Isaac Lab, and a task-and-motion planner generates long-horizon demonstrations without teleoperation, with oracle planners and controllers that show whether high-level planning or low-level execution caused a failure. Trained only in simulation, the ManiFlow low-level policy succeeded on 80.0% of atomic skills in sim and 81.1% on a physical Franka FR3, and a Molmo 2 plus ManiFlow hierarchy completed 55.5% of simulated and 50.0% (15 of 30) of real three-block spelling tasks. In agent experiments, Claude Opus 5.5 driving ToyBlocks skills through an MCP server solved 30 of 30 classical stacking tasks, and skill descriptions the agent wrote after running its own experiments raised success from 24% with default descriptions to 63%, versus 36% with descriptions written by a human expert; among world models tested, Wan 2.2 best predicted block motion and outcomes. Ai2 is releasing the simulation assets, generators, generated demonstrations and two trained models. Source
2. Open ASR Leaderboard Adds Private Egyptian and Nigerian English Test Sets From Liva AI and Reports Accent Groups
Hugging Face added two held-out conversational speech splits from Liva AI to the Open ASR Leaderboard: 1.34 hours of Egyptian-accented and 1.55 hours of Nigerian-accented spontaneous English, transcribed with punctuation, casing and disfluencies. They join the 11 private Appen and DataOceanAI splits across five accents added earlier and a Voice Arena split for Indian-accented English, which the maintainers keep private so models cannot be tuned on the test data. With the new accents, the leaderboard now reports US versus non-US results and, following Kachru’s Three Circles model, Inner Circle accents (US, UK, Australia, Canada) versus Outer and Expanding Circle accents (India, Nigeria, Egypt), with finer splits planned as more accents are added. Source