Tactile Grounding Gets Closer, Benchmarks Get Serious, and Policy Design Gets More Principled

Research

MVT: A Benchmark for Multi-View Tactile Understanding

This paper introduces a benchmark for evaluating how well models understand touch from multiple viewpoints. For builders, this is a concrete tool to measure progress on a core robotics problem: making sense of contact.

SkillDiffuser: Diffusion Policy for Learning Long-Horizon Skills via Skill Decomposition

This work breaks down complex, long-horizon tasks into learnable skill primitives using diffusion models. It's a practical approach to making long-horizon robot behavior more tractable and reusable.

Curriculum Learning for Robot Manipulation with Sparse Rewards

The paper presents a curriculum method for sparse-reward manipulation tasks, a classic challenge. Getting robots to learn with minimal reward engineering is a high-leverage goal for reducing deployment cost.

Adaptive Skill Acquisition for Long-Horizon Tasks

This framework lets robots adaptively learn and reuse skills for long tasks, reducing the need for complete retraining. It's a step toward robots that can build on past experience more efficiently.

From Language to Action: Grounding Instructions with Visual Affordances

This research grounds language commands in visual affordances for more reliable action generation. For anyone building on language models for robot control, this tightens the link between words and physical action.

Tools

GenSim2: Scaling Up Robot Simulation with Automatic Scene and Task Generation

GenSim2 automates the creation of simulation environments and tasks, which is a massive leverage point for training data. If you're spending engineer time hand-crafting sim setups, this points toward a much faster path.

RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots

RoboCasa is a new large-scale simulation platform for training generalist robots in home environments. This is exactly the kind of scalable, realistic data pipeline that could accelerate sim-to-real transfer significantly.