PhysicalMamba: State Space Models for Physical Planning and Robotic Task Execution

PhysicalMamba is a lightweight State Space Model (SSM) architecture designed for physical reasoning, robotic planning, and embodied artificial intelligence systems. This work investigates whether selective state-space architectures can serve as an efficient alternative to Transformer-based models for robotic task execution under constrained computational budgets. We introduce PhysicalMamba-16M, a compact Mamba-inspired language model trained from scratch and evaluated against a parameter-matched TinyTransformer baseline. The study includes large-scale language model pretraining on FineWeb-Edu, supervised fine-tuning on the PhysicalRobotBench-v1 benchmark, and downstream evaluation across robotic planning, memory recall, action execution, and safety-constrained reasoning tasks. Experimental results demonstrate that PhysicalMamba achieves superior downstream robotic reasoning performance despite having a comparable parameter count. PhysicalMamba achieved 21.5% Exact Match, 53.7% Action Accuracy, 100% Safety Constraint Satisfaction, and 53.7% Plan Consistency, outperforming the TinyTransformer baseline on most planning-oriented metrics. The repository contains the complete PhysicalMamba implementation, training pipelines, evaluation framework, benchmark datasets, experiment logs, visualization figures, and research paper source files. This work contributes evidence that State Space Models are a promising architectural direction for future Physical AI systems, robotic foundation models, and embodied agents operating under memory and compute constraints.