RLE-Bench tests coding agents on 51 robot-learning tasks: GPT-6 Astra and Claude Fable 5.1 lead, while mechanical design defeats every model
The authors introduce RLE-Bench, which organizes 51 robot-learning tasks into four workflows—interactive control, policy learning, perception and estimation, and mechanical design—aggregates task-specific metrics into an RLE Index, and evaluates 11 model–harness combinations, finding that GPT-6 Astra and Claude Fable 5.1 lead by a substantial margin, that policy development shows the smallest gaps, and that no model consistently makes sound mechanical design decisions.