Zero-Shot Grasping, Humanoid MoCap, and VLM Safety

Research

AMASS is All You Need: Learning Whole-Body Humanoid Control

Training whole-body control from human MoCap without tedious per-task reward design accelerates the movement of humanoids from lab demos into unstructured, real-world spaces. This massive dataset and framework turn a painful engineering problem into a scalable data pipeline.

Sim2Real Object-Centric Grasping: Zero-Shot Adaptation Framework

This framework eliminates the zero-shot cold start problem so you stop wasting cycles calibrating for new objects. For warehouse robotics or bin-picking, it means deploying a gripper on a Monday morning and having it actually work before lunch.

Tools

Closing the Perception Gap with Integrated VLMs

LT3 makes the dangerous blind spot between vision and language models disappear. If you are tired of your robots hallucinating collisions in dynamic environments, this architecture provides the tactical perception stack you've been missing.

Official Release: SmolVLA Reasoning Model Integration with LeRobot

The ecosystem is finally moving away from brittle, script-based teleoperation toward resilient, multi-step reasoning. The integration with LeRobot facilitates rapid prototyping, allowing you to go from a vague idea to a deployed function much faster than before.

News

Figure AI Leaks Details on 'SpideySense' Synthetic Data Generation

The industry is finally hitting diminishing returns on purely internet-based training, forcing a pivot toward synthetic manipulation data. Figure generating a massive synthetic dataset signals that high-quality robotic training data is now the primary bottleneck for scaling foundation models.

Analysis

Spot's New 'Spatial AI' Data Collection Module (Boston Dynamics)

Spot is quietly becoming the world’s most effective data collection platform, bypassing the slow cycle of human teleoperation to gather high-fidelity 3D maps. This shift from 'robot as worker' to 'robot as data miner' is how you build spatial reasoning models that actually work.