Anchor 1 is a humanoid whole-body controller that reduces tracking error by 30–80% and adds a benchmark for precision, stability, and reachability. The team will release expanded terrain capabilities, model weights, deployment code, and teleoperation tools by the end of 2026.
Key points
30–80% reduction in tracking error reported.
Release of terrain capabilities, model weights, deployment code, and teleoperation tools planned for end of 2026.
Sources
- Original post
- Anchor 1: Precise and Stable Humanoid Control for Whole-Body Manipulation - Linked page
- arxiv.org - Linked in the post
VerneRobotics V0.5 released, leads LIBERO benchmark
VerneRobotics V0.5 is an open-weight model for industrial robots that predicts video frames and robot actions. It ranks as the strongest open-weight model from a US company on the LIBERO and RoboTwin 2.0 benchmarks, outperforming Physical Intelligence’s π0.5 in cluttered scenes.
Key points
Release: V0.5 open-weight model predicts next video frames and robot actions.
Benchmark: tops LIBERO and RoboTwin 2.0, beating π0.5 on cluttered, randomized scenes.
Sources
- Original post
- huggingface.co - Linked in the post
- github.com - Linked in the post
- 22astronauts.com - Linked in the post
OpenAI AI Model Discovers 948 Linux Kernel Vulnerabilities
OpenAI AI Model discovers 948 Linux kernel local privilege escalation vulnerabilities. The security model under development found 2,308 Linux vulnerabilities that allow an attacker with only local privileges to gain root access. The investigation cost $1.2 million (about 190 million yen).
Key points
2,308 vulnerabilities allow local privileges to gain root.
948 distinct local privilege escalation vulnerabilities identified in the Linux kernel.
Investigation cost $1.2 million (about 190 million yen).
Sources
Reinforcement Learning Enables Scaling Generalist Robot Policy Improvement
Reinforcement Learning Enables Scaling Generalist Robot Policy Improvement shows that post-training on 144 diverse tasks with a critic yields up to +58 percentage points improvement on unseen scenes and +39 pp on unseen objects. The work compares critic-guided methods (BON, OGPO) with filtered supervised fine-tuning and highlights task diversity as key.
Key points
Scaling to 144 training tasks gives up to +58 pp gain on unseen scenes.
The same scaling yields up to +39 pp gain on unseen objects.
Sources
Fern launch – high-fidelity robotics simulation platform
Fern by Ishiki Labs, backed by Y Combinator, launches a high-fidelity robotics simulation platform that builds simulations from only three teleop episodes. The platform lets users run thousands of parallel rollouts via an API, with a customer reporting over 900k simulations.
Key points
Launch: Fern by Ishiki Labs, backed by Y Combinator.
Customer ran 900k+ simulations on Fern.
Sources
- Original post
- fern.bot - Linked in the post
Origin Unknown launch – new illegal gold tracking platform
Origin Unknown is a new platform for tracking illegal Venezuelan gold. It was developed with Harvard University researchers and SOS Orinoco. The launch was announced on X with a link to the site.
Key points
Launch: Origin Unknown platform announced
Collaboration: Developed with Harvard University researchers and SOS Orinoco
Sources
Proton communication protocol presented at ROSCon2026
Proton is a lightweight, transport-agnostic communication protocol presented at ROSCon2026. It is intended to bridge custom embedded hardware with modern ROS 2-based systems.
Key points
Proton is lightweight and transport-agnostic
It bridges custom embedded hardware with ROS 2 systems
Sources
- Original post
- bit.ly - Linked in the post
WidowX provides tools to train manipulation skills
WidowX provides tools to train and improve manipulation skills for physical AI. Teleoperation can guide robot arms, collect demos and build datasets for robot learning.
Key points
Tool: WidowX provides tools to train & improve manipulation skills.
Method: Teleoperation can guide arms through physical tasks, collect demos & build datasets.
