Web console

One browser console, six tabs. Which tabs are active is decided by the config passed at launch.

Tab	Role
DEBUG	Run a policy in sim or on the real robot.
MANUAL	Drive a real robot from per-joint sliders.
COLLECT	Teleoperate to record training data.
REPLAY	Open-loop playback of a recorded episode.
EVAL	Score model runs against a benchmark.
RESULT	Browse and compare scored runs across models.

Launching the console

Launch eva with --config pointing at a file under configs/.

bash

# --web-port defaults to 8080; override only if the port is taken.
eva --config configs/01_deploy/dual_agilex_AGILEX/openpi_qpos.py --web-port 8080
# open the URL printed on stdout, e.g. http://localhost:8080

Configs are grouped by intent, then by robot:

Deploy — configs/01_deploy/<robot>/.
Collect — configs/02_collection/ (enables COLLECT).
Evaluate — configs/03_evaluation/ (enables EVAL/RESULT and opens on EVAL).

See Add a Robot for adapting a config to your hardware.

Common failures. eva: command not found → environment not activated. Port busy → pass a different --web-port. Config-not-found → check the path under configs/01_deploy/.

How the console talks to the robot

The console is a remote, not the runtime. The control program already drives the robot; browser buttons dispatch commands (setup, run, stop, switch tab) to it.

Live state refreshed several times per second:

current qpos and the list of available cameras;
3D view of the robot, rendered from its URDF;
per-camera video feeds;
status line: robot/policy connectivity, current step, inference time (ms).

If the URDF is missing or unreadable, the 3D panel is hidden and a warning is logged; everything else continues. Point the config at a valid URDF to restore it (see Add a Robot).

Collection

The tabs aren't isolated features — they fall onto three flows: collection, deployment, and evaluation. Switching tabs resets the tab to a clean state; the shared 3D view and camera strip move with the active tab. Collection is where you verify the hardware and capture demonstration data.

Tab	What it does
MANUAL	Per-joint sliders against a real robot: STAGE a target pose, SEND TO REAL, or HOME. Solid pose = live state, ghost = staged target. See MANUAL.
COLLECT	Teleop-driven recording: pick a task, START/STOP, watch the conversion to dataset format, replay an episode to mark PASS/FAIL with a note. See COLLECT.
REPLAY	Open-loop replay over a dataset: load a folder and an episode index, scrub frames against cameras and 3D view, optionally re-run the policy. Per-episode PASS/FAIL and language notes. See REPLAY.

Deployment

Connect the policy and run inference — in simulation preview or on the real robot.

Tab	What it does
DEBUG	Default tab. Pick a prompt and a mode — `real`, `sim`, `step` (one chunk per advance), or `manual` — then SETUP, RUN / STOP / RESET. Inference rate and action publish rate are tunable. See DEBUG.

Evaluation

Score checkpoints on the real robot and browse the results.

Tab	What it does
EVAL	Scoring flow MODEL → TASK → TRIAL → RUN → SCORE (milestones plus note). Each run is recorded with its score. See EVAL.
RESULT	Searchable/sortable table of scored runs, side-by-side model comparison, replay of any run's cameras, 3D view, and charts. See RESULT.

Why some tabs are greyed out

All six buttons always render; activation is decided at startup from the launched config.

COLLECT is disabled unless the config defines a collection schema. configs/02_collection/ sets this up.
EVAL / RESULT activate when the config defines an eval block. configs/03_evaluation/ sets this up; when present, the console opens on EVAL. Without it, EVAL/RESULT remain a read-only viewer over prior results.
MANUAL always renders, but SEND TO REAL only reaches hardware over a live transport (zmq, ros1, ros2). A dataset-backed config has no live robot; the UI flags this.

Tab availability is config-driven. A plain deploy config yields DEBUG + MANUAL + REPLAY, with COLLECT disabled and EVAL/RESULT read-only. See Configuration.

Connection & liveness

The status bar's online flag means a message arrived recently — not just that a socket opened. For zmq/ros1/ros2, the bar stays offline until the first message and flips back to offline after 2 s of silence. Dataset/replay sources report whatever the config declares.

A separate indicator tracks policy connectivity. The console also pings the running process on every refresh so it can detect a closed browser.

Stays offline. Check power, link, the eva process for errors, and that the config's transport matches how the robot is wired.