What's Actually Working in Robot Learning: World Models, VLAs, and Where the Failures Hide

Paper

High-Frequency VLA Evaluation: When Speed Becomes the Control Problem

High-frequency evaluation shows VLAs can close the sim-to-real gap when you actually run them fast enough. If your VLA runs at 5Hz, you're already smearing latency into the control loop. Worth benchmarking your stack honestly.

Video-Conditioned World Models Reduce Reward Specification in Robot Learning

Video prediction as a world model weakens the need for reward engineering at evaluation time. Useful for open-ended manipulation but still fragile outside narrow task families.

Hierarchical Sub-Policy Decomposition Fails Under Non-Stationary Task Distributions

A straightforward negative result: hierarchical policies collapse when you decompose the wrong way. The lesson is that your compositional inductive bias matters more than your architecture sophistication.

LLM-Guided Multi-Agent Coordination Under Partial Observability

LLMs can generate plausible-looking multi-agent plans that destroy real-world rewards because they optimize for coherence, not feasibility. Fine if you have a verifier; dangerous if you trust the raw output.

Locomotion Policy Transfer Across Morphologies: The Sim-Reality Fidelity Gap Widens

Cross-morphology transfer works in tight simulation but fails hard on real contact dynamics. Highlighting that your simulator's contact model is the real bottleneck, not your policy architecture.

Conference

ROSCon 2025: The ROS Community Infrastructure Hashing Out Production Realities

ROS 2 is still the default plumbing for anyone shipping real robots. ROSCon is worth watching if you're deciding on middleware for your next commercial fleet.