SPAR preprint reports 1.98× faster inference for compressed robot model

Researchers report a 1.98× inference speedup for a compressed robot action model in a preprint introducing SPAR, a method that removes model parameters and then uses reinforcement learning to recover manipulation performance. Applied to π₀.₅, it reduces the full model from 3.37B to 1.12B parameters and its language backbone from 2.34B to 0.09B. SPAR removes dimensions within the language backbone while tuning the model. Subsequent reinforcement learning raises the compressed model's average success across four LIBERO benchmark suites from 84.55% to 93.45%, an 8.90-percentage-point gain. That success-rate comparison measures recovery after compression; the reported inference speedup compares the compressed policy with the unpruned model.