EmbodiedJev

Neither. It is MuJoCo simulation with a Franka Panda and three built-in tasks: carry into tray, stack a block, and carry over an obstacle. The default rule baseline runs with no API key, GPU, or model weights, so the full observe-decide-execute loop works before you spend anything.

No. Jev receives structured text state plus RGB-D detector coordinates. Raw camera RGB is only sent in the direct-image stepwise modes for models that accept images. The README states plainly that Jev can read coordinates or vision detections but cannot receive camera pictures directly.

They are a small development comparison, not a suite. On the LIBERO drawer-close task with dual RGB-D cameras, pure GPT-6 cost about $2.15 over 610 seconds versus about $0.39 over 320 seconds for the GPT-6 plus Jev mix. The README says this is one initial state and cannot be extrapolated to a full LIBERO success rate.

Messy so far. Two hierarchical XYZ carry runs both reached a terminal state at 89 and 88 steps with 178 and 176 real requests, but both lost the grasp on the way down and only completed after the block landed in the tray. The README calls this out as not yet stable precise placement.

Python 3.11 or newer, Node.js 22.12 or newer and Git. Create a venv, run python -m pip install -e ., then npm ci and npm run build, and start embodied-jev serve --port 8090 and open http://127.0.0.1:8090. Leave the decision model on the rule baseline for the first run: that path makes zero model calls. The original code is MIT with the Franka Panda assets under Apache-2.0.

Preset skills have the model pick a phase and then a program-generated skill, and are skipped entirely when only one phase is feasible. Stepwise XYZ offers candidates of 40, 10 and 2 mm along each axis plus open, close and hold on the gripper, with the program owning IK, joint control and safety checks. Hierarchical XYZ has the model choose a subgoal first, then the XYZ direction and gripper together, in 2 to 12 mm steps per axis.

Yes. Model comparison runs two or three configurations on one task, seed and action budget, and aligns replays by simulation time rather than wall clock, so a slow model is not flattered by playback speed. You choose sequential or two-way parallel, unconnected models produce no simulated scores, and exports carry each configuration, decision and trace. Replays live in memory and there is no import path for past files yet.