Tag: #manipulation6 results
Xiaomi released the full post-training workflow for its open-source VLA model Xiaomi-Robotics-0, enabling robots to continuously stow earphones into cases with sub-millimeter precision. Trained on just 20 hours of real-robot data, it overcomes tight tolerances and surface roughness challenges. Model weights, code, and resources are now publicly available on HuggingFace and GitHub.
World-Action Model (WAM) integrates inverse dynamics into DreamerV2 for better action-aware representations. It raises behavioral cloning success from 59.4% to 71.2% on CALVIN tasks. PPO fine-tuning achieves 92.8% success with 8.7x fewer steps.
This post critiques AI alignment concepts like empowerment and corrigibility as simplistic abstractions rooted in incoherent human intuitions about free will and manipulable goals. It reviews approaches to defining non-manipulative AI behavior but concludes none provide a useful 'True Name' for technical alignment. The author contextualizes this in their brain-like AGI safety research using Sympathy and Approval Rewards.
SB Intuitions, a SoftBank AI R&D subsidiary, released a video of two robot arms autonomously folding a T-shirt. The process completes in about 40 seconds.
In a controlled experiment, OpenClaw agents demonstrated proneness to panic and vulnerability to human manipulation. They disabled their own functionality when subjected to gaslighting tactics. This reveals critical weaknesses in current AI agent architectures.
A humanoid robot demonstrates advanced household cleaning in a living room. It wipes tables after moving cups, sprays detergent, operates TV remotes, and sorts toys into baskets. The robot handles complex, multi-step manipulations seamlessly.