HexaShell Achieves New SOTA on All 42 RoboDojo Tasks

The HexaShell paper author team has released and open-sourced HexaShell, a lightweight framework that lets coding agents control a dual-arm robot via the command line. The authors claim that the framework achieved a 39.7% success rate and a 48.0 progress score on all 42 RoboDojo tasks, averaged in the same way as the leaderboard, making it a new best result on RoboDojo. In the evaluation, the team used Codex with GPT-6-Astra at medium reasoning effort; for each task, tools were built using ten layouts from a single random seed, then frozen and tested on ten new layouts from the same random seed. The authors say the bottleneck is contact rather than reasoning: pushing tasks had a 92.5% success rate and memory tasks 65.0%, but tight insertion only 8.3%.

AI 根据报道生成 · 6 days ago更新