Skip to content
Drix10 Blog

W2A achieves 95.9% success and 21× faster inference

, 6 items in AI and Robotics Applications, 3 min read

In this digest (6 items)

W2A achieves 95.9% average success on LIBERO and runs about 21× faster than the representative WAM baseline. It distills intermediate denoising representations from a video diffusion model into an action model and discards the video model at test time.

Key points

  • Success rate: 95.9% average on LIBERO

  • Inference speed: about 21× faster than the representative WAM baseline

Sources

Claude Opus 5.5 edits π₀.₅ weights, 30% less task time

Claude Opus 5.5 edits π₀.₅’s weights, achieving ~30% less task time and 100% success in bottle-placement trials. The edit was done without backpropagation or production inference-code changes in about 111 minutes on a single machine.

Key points

  • Edit result: ~30% less task time, 100% success in bottle-placement trials.

  • Edit performed in ~111 minutes on one machine, no backpropagation, no production inference-code changes.

Sources

Agentic robot control runs 2.5 h after 30 s recharge

Agentic robot control runs autonomously for 2.5 hours after a 30-second material recharge. It assembles one gadget every 7 minutes and stops when parts run out.

Key points

  • Runtime: 2.5 hours autonomous after 30 seconds recharge

  • Production: 1 gadget per 7 minutes, stops due to parts shortage

Sources

Reflex humanoid control catches objects without falling, 1.6M rendered steps

Reflex, a Visually Grounded Reactive Humanoid Control system, lets a Unitree G1 catch boxes thrown by people using only an onboard RGB-D camera. The robot repositions its feet and torso, forms a support with its forearms and chest, absorbs momentum and stays upright. Training runs in three stages in Isaac Lab and requires about 1.6 million rendered steps.

Key points

  • Reflex enables a Unitree G1 to catch boxes with an RGB-D camera, repositioning feet and torso and using forearms and chest for support.

  • Training consists of three Isaac Lab stages and uses roughly 1.6 million rendered steps, with a latent recovered from 25 frames of egocentric RGB-D history.

Sources

AIR humanoid launch by EmbodiedAI in Switzerland

AIR is a Swiss humanoid announced by EmbodiedAI for safe, reliable industrial automation. The post also introduced the @EmbodiedAI_ch account.

Key points

  • Launch: AIR, Swiss humanoid for safe, reliable industrial automation.

  • Account: @EmbodiedAI_ch announced.

Sources

Railway IP addresses blocked by network providers

Railway reports that network providers and security services have blocked Railway IP addresses, preventing access to Railway-hosted applications. The issue is confirmed and its status is Identified.

Key points

  • External network blocks affecting access to Railway

  • Status: Identified

Sources