What our run established
Our run on October 7, 2026, at commit f08de2e recorded npm installation in 24 seconds and a front-end build in 10 seconds; both succeeded, and tests were skipped.
These are observations from our sandbox. The matching harness stores command outcomes separately from retained log tails: both commands returned exit code 0, neither timed out, and the test step records a skip because there was no default script. A skipped test step is not a failed suite. The successful build establishes that the front-end assets compiled in this environment.
We did not install the Python package, start the service, open the browser interface, execute a MuJoCo task, load MiniCPM or call a model provider. The manifest defines separate UI and site test commands, and the technical guide documents Python tests; our run did not execute them. The raw build output also contains a chunk-size warning despite its successful exit.
The MrKeyoor testing methodology explains how command outcomes and measurement limits reach these reviews. Its supplied October 31 description provides provenance context, not proof that the same methodology version governed our October 7 run. This correction includes no new execution.
What the model controls
EmbodiedJev is a browser workbench that uses MuJoCo to simulate a Franka Panda arm. Its technical guide assigns display and controls to the browser, with physics, model requests and experiment records handled by Python.
The built-in Panda tasks cover moving a block into a tray, stacking it on a fixed block and carrying it over an obstacle. Models select skills or short XYZ and gripper actions, depending on the mode. Program logic handles action conversion, inverse kinematics, execution checks and physical success conditions. The guide says collision monitoring is incomplete and the end-effector orientation is fixed. These are documented capabilities and limits, not behavior we reproduced.
Which comparisons are meaningful
Simulator coordinates, locally detected RGB-D coordinates and raw camera images give models different information. The README separates these inputs and distinguishes prepared-skill selection from incremental movement decisions. Keep those protocols separate when interpreting results: selecting a skill from coordinates answers a different question from choosing movements from images.
The documented comparison interface runs configurations with a shared task and seed and aligns trajectories by simulation time. Replay speed does not establish inference latency because model waiting time is separate. Preserve input mode, controller, seed, action budget and actual elapsed time alongside exported results. The authors' experiment results are not our passing-test counts or independent reproductions.
What to verify after building
Start with the rule baseline, which the documentation says needs no API key, GPU or model weights after installation. Start the Python service, open the local browser page and complete a baseline task before adding a model connection. That checks the application path our npm run left untested.
The supplied Chinese README and technical guide describe setup, protocols and limitations in detail. Windows launch commands are provided, but the README identifies macOS and Linux CI as validated environments. MiniCPM adds optional Python dependencies, about 5 GB of initial weights and further runtime memory. Cloud routes require provider credentials; connection tests and experiments make real requests and may incur charges.
Cloud calls send task state, and direct-image mode also sends enabled cameras' RGB frames. Localhost binding does not prevent that transmission. Decide whether the observations may leave the machine before testing a provider connection.
Why the audit result remains unresolved
The stored audit summary records zero advisories, while the raw installation tail reports one high-severity vulnerability. We cannot reconcile those outputs from the supplied record and do not describe the installation as vulnerability-free. Scan coverage is unknown because the record has no scannedPackages field.
The attached stored parser summaries also record 26 dependency packages and a 148 MB footprint, while the installation tail reports a different added-package count. We preserve the stored values without treating the tail as a replacement measurement. No complete-log reconciliation was supplied, and no test total exists because tests were skipped.
When another project fits
robosuite is a MuJoCo simulation framework for robot learning with benchmark environments for reproducible research. Its primary repository describes modular APIs for environments, robot embodiments and controllers. Choose it when building those simulation components is your main task.
Meta-World is a benchmark for developing and evaluating multi-task and meta reinforcement learning algorithms for robotic manipulation. Its primary repository documents the benchmark environments and Gymnasium interface. Choose it for that evaluation role. EmbodiedJev also reports Meta-World experiments, so the benchmark can complement the workbench rather than replace its browser interface.
How to read the ratings
The editorial rubric runs from 1 to 5: 1 indicates substantial obstacles, 2 substantial gaps with a usable path, 3 useful foundations with material limits, 4 detailed support with a bounded shortcoming, and 5 strong evidence across the dimension. These ratings are judgments about the supplied evidence, not performance measurements.
Setup ease is 3 because npm installation and build succeeded but the Python service and simulator remain unverified. Documentation is 4 for detailed setup and limitation explanations, with Chinese onboarding as a constraint. Community is 3 only for documented contribution and reproduction paths; it does not establish adoption or maintainer responsiveness. Maturity is 3 for detected workflows and test directories, qualified by skipped tests and the narrow build check.

