Adaptive Morphologies, Efficient RL, and Manipulation Benchmarks Get Sharper

Research

Self-Modifying Robot Morphologies and Gaits via Learned Optimization

Teaches robots to modify their own structure and gait on the fly. High-leverage idea: instead of training separate controllers for every new shape, a single network adapts. If this scales, it cuts the cost of multi-morphology deployment significantly.

Decoupled RL: Accelerating Reinforcement Learning for Continuous Control

Demonstrates faster sample-efficient reinforcement learning by decoupling reward signals from temporal-difference targets. Speeds up the training loop substantially for continuous control tasks like robotics simulation benchmarks. Worth benchmarking against your current pipeline.

An Object-Centric Benchmark for Everyday Robotic Manipulation

Introduces tighter ground-truth metrics for evaluating robotic manipulation by grounding task success in physical object state changes. Fixes a common blind spot where simulated performance scores hide failed real-world outcomes. Useful for builders benchmarking new architectures.

Generating Spatial Representations for Robotic Navigation from Sparse Views

Combines structure-from-motion with generative modeling to create generalizable spatial representations from sparse viewpoints, no dense map needed. Reduces sensor overhead for navigation-heavy robotics systems. Worth watching for anyone building mobile or autonomous platforms.

Discussion

Why Humanoid Robotics Is (Finally) Having Its Moment

Podcast conversation unpacking recent progress in humanoid control, real-world deployment bottlenecks, and the economics of scale. Provides useful context for builders deciding whether to bet on humanoids or specialized arms.