RoboJEV

No. The independent rule baseline scored 48 of 50 on the same physical scenes with the same success checks, and it is never used as a fallback for JEV. The gaps are stack on a pedestal at 8/10 versus 10/10 and gate pick and place at 5/10 versus 8/10; the other three tasks tied at 10/10. The claim is narrower, that task-conditioned control works from privileged state.

Not on their own, and the project says so. Each result carries a Wilson 95% interval, so even 10/10 leaves 72.2% to 100%, and gate pick and place spans 23.7% to 76.3%. Ten fixed seeds per task and policy is a small sample rather than a claim of general-purpose manipulation, and every completed failure stays in the denominator across the two frozen campaigns.

It sees structured simulator state, not images, which is the honest framing. JEV gets TCP and object poses, velocity, jaw width, contact measurements, geometric relationships, and recent actions, with no camera input and no evaluator history; task instructions and geometric helpers are engineered explicitly. The README also rules out learned vision, robot training, arbitrary manipulation, and real-hardware deployment.

Nothing executes. Malformed or inconsistent responses are validated and produce no action, and physics pauses during API calls so nothing moves mid-request. Each two-stage decision costs at least two API requests. Transport handling is a separate story worth knowing: the README records all 113 errors recovered during one campaign and 38 HTTP 402 responses logged as account blocks.

Linux with Python 3.11 is the tested platform. Create the conda environment from environment.yml, install the package with the test and video extras, then copy .env.example to .env and set TYPESAFE_API_KEY. CPU physics and the rule policy need no GPU or API key; only recording needs a graphics backend.

No improvement was observed. Single gate: original 3/10 versus full geometry 2/10, with both failing on 7 of 10 paired scenes; double gate failed 10/10 either way. Exact McNemar two-sided p=1.00 per task, and ten pairs are too few to establish equivalence or harm.

Nonzero axis choices form a normalized 1 cm translation in robot-base coordinates, reduced to 4 mm while a grasped peg is within 25 mm of the socket axis; the gripper command is open, hold or close. Execution is Jacobian damped-least-squares IK with joint position control at 2 ms physics steps.