REPLAY

Plays a recorded episode back through the dataset transport — no policy server, run continuously or release chunk-by-chunk.

Overview

REPLAY is one of the console's six tabs (DEBUG · MANUAL · COLLECT · REPLAY · EVAL · RESULT). DEBUG asks a live policy for the next action; REPLAY reads actions from a recording on disk and re-publishes them in order. Anything COLLECT or an inference run records can be replayed — both save the same LeRobot v2.1 format. See Data collection for the on-disk layout.

Loading an episode

Launch against the dataset-replay preset (see Arming REPLAY from config):

bash

eva --config configs/00_openloop/dual_agilex_AGILEX_openloop.py --web-port 8080

Open http://localhost:8080 and click REPLAY. Fill in the left-panel inputs; the console mounts that episode as a read-only source for the session — no relaunch needed.

Input	What to enter
Dataset directory	Path to the recording's root folder — the one holding `meta/`, `data/`, `videos/`. Example: `examples/agilex_dataset/`. The console reads this first to learn the episode count and columns.
Episode	Which run to play, 0-indexed. A 5-episode dataset accepts 0–4. The count is shown after inspection.
State / action columns	Which saved columns supply state and action. Leave blank to infer from the dataset.
Action mode	Whether the recorded action is `qpos` or `eef`. EEF actions are resolved through the same IK step a live run uses.

Once loaded, the playback speed is adjustable and a seek bar scrubs to any frame, freezing the 3D arm there. Current and total frame counts appear in the status feed.

Playing back: continuous or step-by-step

Like DEBUG, REPLAY offers two modes: continuous playback, or pause before every chunk and release one at a time. The only difference from DEBUG is the chunk source — disk, not a live policy.

Control	What it does
RUN / STOP / RESET	Play continuously, halt, or return the arm to its start pose.
SIM ↻	Advance one chunk on the simulated arm only — the real robot does not move.
REAL ▶	Send the previewed chunk to the real robot.

SIM-then-REAL is the same staged release flow as DEBUG — see DEBUG → STEP.

The live view

The right side is the shared 3D view. During playback:

3D arm — driven by the joint angles of the chunk in flight.
Camera strip — the recorded camera videos saved with the data.
ACTION / STATE charts — per-axis graph of the chunk in flight; on stop, a scrub bar unlocks the history.

Full live-view mechanics: DEBUG → The live stage.

Arming REPLAY from config

REPLAY can be armed entirely from config so actions are read from disk on launch. EVA ships dataset-replay presets per robot at configs/00_openloop/<robot>_openloop.py. These set the transport to read from a saved dataset — no robot, no ROS, no policy server.

Copy the preset matching your robot (dual-arm AGILEX: configs/00_openloop/dual_agilex_AGILEX_openloop.py) and edit the two fields pointing it at your recording. Transport keys:

Key (under `transport`)	Purpose
`type`	`'dataset'`. Leave as-is.
`dataset_dir`	Required — path to the recording's root folder.
`episode_id`	Which episode to play, 0-indexed. Defaults to 0.
`dataset_keys`	Maps saved columns to state/action/video. Defaults match EVA recordings; an empty `video_keys` falls back to `observation.images.{camera}`.

python

transport = dict(
    type='dataset',
    dataset_dir='<path/to/your/dataset>',  # ← required.
    episode_id=0,
)

Launch:

bash

eva --config <path/to/your_openloop.py> --web-port 8080

vs. the replay policy backend. Use the dataset-replay preset for plain playback. Use the replay policy backend only when the recording must flow through the normal policy machinery; see Backends.

What you can replay

Any LeRobot v2.1 recording EVA produces:

Source	Contents
COLLECT episodes	Teleop recordings — measured state paired with the demonstrated action.
Inference runs	One episode per run — state paired with the action the policy sent.

A sample dataset ships at examples/agilex_dataset/ (dual-arm AGILEX, three cameras). Point Dataset directory or dataset_dir at it. On-disk layout: Data collection.