RoboticsLiberAIConfirmed
Story 15 of 17 in the Wednesday 4 November editionLiberAI unveils Liber-0 robotics World Action Model trained on 2.2M human video hours
Paired with LiberSense motion-capture gloves and a Unified Embodiment Representation, Liber-0 reaches a 36.0% zero-shot success rate across 60 GPT-6 manipulation tasks.
LiberAI introduced its Liber-0 World Action Model on 4 November 2026, pretraining a Unified Embodiment Representation on over 2.2 million hours of human video anchored by hand-trajectory recordings from its LiberSense Glove system and humanoid LEO hardware, achieving a 36.0% task-level zero-shot success rate across 600 trials on 60 unseen manipulation tasks proposed by GPT-6 without task-specific robot demonstrations; weights were not released.
Sources
- PrimaryLiber-0 technical report on emergent zero-shot manipulationliberai.com
The rest of the day
Read the whole day01CreativeOpenAI
OpenAI rolls out ChatGPT Images 2.6 with animated GIFs and two new API image models
Images 2.6 generates visuals 50% faster across ChatGPT and Codex while adding GPT-Image-2.6 Flare and Sunburst to the API with SVG export and markup-guided edits.
02ModelTencent
Tencent launches multimodal Hy4 model with native computer use and 3D engine training
Hy4 adds visual reasoning alongside CLI and GUI control, trained directly inside Blender, CAD, Unreal Engine, Unity and Godot environments.
03Open sourceMoonshot AI
Moonshot AI releases open-weight Kimi K3.1 Pro with 2.8T parameters and 1M context
Activating 104B parameters per token with native MXFP4 weights, the multimodal MoE model leads Opus 5.5 and GPT-6 Astra on SWE Marathon v1.1.
04ResearchOpenAI
OpenAI’s autonomous cyber pipeline uncovers and patches 948 Linux kernel root flaws
Powered by an upcoming cyber model, a five-stage VM validation and exploit pipeline generated working privilege-escalation exploits at $1,279 per distinct vulnerability.
Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X







