Energy-Based Policy Orchestration (EPO) is introduced, reducing policy switches by 7× and providing differentiable energy scores for policy-goal pairs. Adding a new skill only requires training one new energy model, without retraining existing policies or a central scheduler.
Key points
7× less frequent policy switches compared with learned MoE routers.
Adding a skill means training one new energy model only.
Sources
- Original post
- Energy-Based Policy Orchestration - Linked page
- yewr.github.io - Linked in the post
- github.com - Linked in the post
Gaussian Splats on Vision Pro generated in 20 seconds
Gaussian Splats viewer on Vision Pro generated in 20 seconds using an RTX 5080. The GS viewer app for Vision Pro is currently in beta.
Key points
RTX 5080 used for TripoSplat generation
GS viewer app for Vision Pro is in beta
Sources
GPU without Linux enables GUI under a second startup
GPU without Linux enables a full GUI in under a second after power-on. AI shaved milliseconds off the process and Slint+Zephyr is noted as fun.
Key points
GPU without Linux: full GUI up in under a second after power-on
AI shaved milliseconds off the startup process
Sources
LoRA adapters training live on Hugging Face
LoRA adapters training is live on Hugging Face, using RL on LFM2.5, LFM3 and Twil-9B models. The runs can be followed live.
Key points
LoRA adapters are being trained on LFM2.5 and LFM3 from liquidai.
LoRA adapters are being trained on the Twil-9B model from thewebAI using reinforcement learning.
Sources
Lorenz Begins Operating Einride Autonomous Trucks
Lorenz begins operating Einride autonomous trucks, with four driverless electric vehicles moving goods between its plant and warehouse. The trucks transport finished goods, raw materials, and empty pallets.
Key points
Vehicles: Four driverless autonomous electric vehicles are in operation.
Cargo: They move finished goods, raw materials, and empty pallets between the production facility and the central warehouse.
Sources
Gaussian Splats generation from image viewable on Vision Pro
Gaussian Splats can be generated from a single image and viewed on Apple Vision Pro, representing a leap forward for spatial computing. The workflow described goes from image to 3DGS to Vision Pro display, noting USDZ struggles with fluffy materials.
Key points
Workflow: image → 3DGS → displayed on Vision Pro
USDZ (3DCG) struggles with fluffy materials, making Gaussian Splats preferable
Sources
Zephyr project released, enables GPU on bare metal
Zephyr project is now public and enables GPU usage on bare metal. libGLES.a is provided that works outside of Linux.
Key points
Zephyr project made public.
libGLES.a works outside of Linux.
Sources
- Original post
- yako.gl - Linked in the post