Embodied Agents Get Memory, Contact-Rich Tasks Get Cleaner Data, and Manipulation Policies Get More Generalizable

EmbodiedRAG: Retrieval-Augmented Memory for Long-Horizon Robot Tasks

Giving robots retrieval-augmented memory lets them reference past experience without retraining, which is a practical unlock for long-horizon tasks. This is the kind of architecture decision that separates one-shot demos from robots that actually accumulate useful context over time.

Contact-Implicit Trajectory Optimization with Learned Residual Dynamics

Contact-rich manipulation is notoriously hard to simulate cleanly, and learned residual dynamics on top of physics priors is a pragmatic middle path. If this holds up, it cuts down the sim-to-real gap for tasks like insertion and assembly without throwing out the physics model entirely.

Cross-Embodiment Policy Transfer via Morphology-Agnostic Representations

Training a policy on one robot body and deploying it on another without retraining is the holy grail for reducing data collection costs. This paper's morphology-agnostic representation approach is a direct attack on one of the biggest bottlenecks in scaling manipulation research.

Real-Time Semantic Scene Graphs for Mobile Robot Navigation in Unstructured Environments

Scene graphs give robots a structured world model that pure pixel-based perception can't easily query, and running them in real time on mobile hardware is the hard part. This closes a gap between rich semantic understanding and actual deployment speed.

Scaling Robot Foundation Models with Heterogeneous Multi-Robot Datasets

Pooling data across different robot platforms is how you get foundation models that generalize, but heterogeneous datasets introduce serious distribution mismatch problems. This work's approach to handling that mismatch is worth tracking if you're building anything that needs to work across hardware variants.

Autonomous Tool Selection and Use in Unstructured Manipulation Tasks

Robots that can autonomously pick the right tool for a task without explicit programming are a step change in flexibility for unstructured environments like warehouses and kitchens. This is early but directionally important for anyone building general-purpose manipulation systems.