John Coogan

Interleaving Model-Based Motion Planning with VLAs Doubles Long-Horizon Manipulation Performance — No Extra Training Required

Current Vision-Language-Action (VLA) models fail on long-horizon mobile manipulation tasks due to compounding early errors and difficulty maintaining goal progress across spatially distributed subtasks — and finetuning on more data does not resolve this. MPVI (Motion Planner/VLA Interleaving) addresses this by integrating classical model-based motion planning with VLAs at inference time, using open-vocabulary object detection, frontier exploration, and VLM-based completion checking with proprioceptive triggers to manage module switching. Evaluated on the BEHAVIOR-1K benchmark, MPVI achieves a 113% improvement in task progress over a top end-to-end VLA baseline without any additional training.

Hybrid MARL + MPC Framework Achieves 100% Success in Multi-Agent Hardware Deployment vs. 60% for Pure Neural Baseline

The paper introduces MA-AC-MPC, a hybrid framework that couples multi-agent actor-critic reinforcement learning with Model Predictive Control (MPC) to address the complementary weaknesses of each approach: MARL handles long-horizon cooperative policy learning from sparse rewards, while MPC enforces short-horizon dynamic feasibility and safety constraints. The combined algorithm is validated in two hardware scenarios — a pursuit-evasion task and a heterogeneous drone/rover cooperative landing task — where it significantly outperforms a comparable MARL-with-MLP baseline. The 100% vs. 60% hardware success rate gap between MA-AC-MPC and MA-AC-MLP on the landing task provides a concrete, quantified advantage for the hybrid architecture in real-world deployment.

Observation of the charmless purely baryonic decay $\mathinner{\mathitΛ^0_b\!\to \mathitΛ p \overline{p}}$

Mazocarta: A Seeded Procedural Deckbuilder for Instrumented Game Development