HexaShell hits 39.7% on RoboDojo, finds contact is the wall

HexaShell, a thin harness that lets a vision-language model drive a two-arm robot by writing and fixing its own code tools, reaches a 39.7% success rate and a 48.0 progress score across all 42 RoboDojo tasks, which the authors of the arXiv preprint call a new state of the art on the benchmark. After each failed attempt, the agent reads the log, finds the cause, and keeps what it learns as code, "as a coding agent does when it fixes a failing test."

The task-type breakdown is where the paper's real claim lives. Pushing reaches 92.5% and memory tasks 65.0%, but tight insertion lands at 8.3%. The authors are blunt about what that pattern means: "The limit is contact, not reasoning."

The harness itself is deliberately minimal. The agent drives both arms through one command line, sees three calibrated RGB cameras, and locates objects using triangulation. "If two cameras see the same point, the point lies where the two camera rays meet," the paper reports; when the head camera is not enough, the robot moves a wrist camera to get a second view.

The protocol freezes what the agent learns before scoring. Tools are built on ten layouts of a single seed per task, then frozen and tested on ten new layouts of the same seed. The underlying model is Codex with GPT-6-Astra at medium reasoning effort.

The failure mode the paper names is specific, and it is not a planning or perception failure in the usual sense. "When the agent fails, it usually sees the failure but cannot make the motion work."