Humanoid Robot Locomotion and Manipulation Advances

The field of humanoid robotics is rapidly advancing with a focus on improving locomotion and manipulation capabilities. Recent developments have centered around creating more stable and adaptable controllers for diverse humanoid embodiments. This has involved the use of hierarchical control architectures, cross-humanoid locomotion pretraining, and sim-to-real policy transfer. Additionally, there has been a push towards enabling humanoid robots to interact with and manipulate various objects in their environment, with advancements in material generalization and physics-based character control. Notable papers in this area include: H-Zero, which introduces a cross-humanoid locomotion pretraining pipeline enabling zero-shot and few-shot transfer to novel humanoid robots. Modality-Augmented Fine-Tuning of Foundation Robot Policies, which demonstrates the effectiveness of modality augmentation in adapting foundation robot policies to diverse humanoid embodiments. Learning Sim-to-Real Humanoid Locomotion in 15 Minutes, which achieves rapid training of humanoid locomotion policies in just 15 minutes with a single RTX 4090 GPU.

Sources

A Hierarchical Framework for Humanoid Locomotion with Supernumerary Limbs

H-Zero: Cross-Humanoid Locomotion Pretraining Enables Few-shot Novel Embodiment Transfer

Opening the Sim-to-Real Door for Humanoid Pixel-to-Action Policy Transfer

Discovering Self-Protective Falling Policy for Humanoid Robot via Deep Reinforcement Learning

Modality-Augmented Fine-Tuning of Foundation Robot Policies for Cross-Embodiment Manipulation on GR1 and G1

$\mathbf{M^3A}$ Policy: Mutable Material Manipulation Augmentation Policy through Photometric Re-rendering

Learning Sim-to-Real Humanoid Locomotion in 15 Minutes

CoatFusion: Controllable Material Coating in Images

SMP: Reusable Score-Matching Motion Priors for Physics-Based Character Control

X-Humanoid: Robotize Human Videos to Generate Humanoid Videos at Scale

Bridging Simulation and Reality: Cross-Domain Transfer with Semantic 2D Gaussian Splatting

From Generated Human Videos to Physically Plausible Robot Trajectories

Built with on top of