The field of robotics and human-computer interaction is witnessing significant advancements in adaptability, precision, and realism. Innovations are particularly notable in the areas of environmental adaptability for robots, where new methods are being developed to enhance robots' ability to interact with objects of varying shapes and hardness without extensive data acquisition. In human motion synthesis, there's a push towards generating more natural and semantically controlled movements that adapt to complex environments, leveraging diffusion models and hierarchical scene reasoning. The identification and estimation of dynamic friction in robotics are also seeing progress through the integration of probabilistic latent variable modeling, improving the precision of control design and friction compensation. Transfer learning from non-human species is emerging as a promising approach to enhance human pose estimation, especially in clinical settings. Co-speech motion generation is achieving new levels of semantic richness and rhythmic consistency, while the generation of interactive motions, such as duet dances, is becoming more realistic through the development of large-scale datasets and diffusion-based frameworks. The restoration of physically plausible 3D human motion from videos, even for high-difficulty motions, is being addressed with innovative plug-and-play approaches. Lastly, the learning of generic skills for humanoid robots through mimicking human data is being facilitated by comprehensive benchmarks, and zero-shot human-scene interaction synthesis is being enabled by integrating video generation with neural human rendering.

Noteworthy Papers

• An Environment-Adaptive Position/Force Control Based on Physical Property Estimation: Introduces a method for generating highly adaptable robot actions with minimal data, significantly improving environmental adaptability.
• SCENIC: Scene-aware Semantic Navigation with Instruction-guided Control: A diffusion model that generates human motion adaptable to dynamic terrains with semantic control through natural language.
• Probabilistic Latent Variable Modeling for Dynamic Friction Identification and Estimation: Proposes a novel approach to friction model identification using latent dynamic states, enhancing precision in robotics.
• Monkey Transfer Learning Can Improve Human Pose Estimation: Demonstrates the effectiveness of transfer learning from macaque monkeys to improve human pose estimation in clinical situations.
• SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis: Offers a method for generating co-speech motions with enhanced semantic richness over a stable base motion.
• InterDance: Reactive 3D Dance Generation with Realistic Duet Interactions: Presents a large-scale dataset and a diffusion-based framework for generating realistic duet dance motions.
• A Plug-and-Play Physical Motion Restoration Approach for In-the-Wild High-Difficulty Motions: Introduces a novel approach to restoring physically plausible 3D human motion from videos, even for challenging motions.
• Mimicking-Bench: A Benchmark for Generalizable Humanoid-Scene Interaction Learning via Human Mimicking: The first comprehensive benchmark for learning humanoid-scene interaction skills through mimicking human data.
• ZeroHSI: Zero-Shot 4D Human-Scene Interaction by Video Generation: Enables zero-shot synthesis of human-scene interactions by leveraging video generation and neural human rendering.