COLLECT

COLLECT records teleoperation episodes to a standard robot-learning dataset, with PASS/FAIL review in the same tab.

Overview

COLLECT is the third of the console's six tabs (DEBUG · MANUAL · COLLECT · REPLAY · EVAL · RESULT). It runs no policy: each episode records the arm's state alongside the operator's commanded action, so the dataset is replayable and trainable in one pass.

Tab is gated by config. COLLECT is enabled only when the launched config carries a collection section (see below). No separate process.

Prerequisites

eva on PATH (see Installation).
Robot powered, leader/follower rig running.
Data source publishing — for ROS, drivers and camera nodes up (e.g. roscore plus arm/camera nodes).
A collection config for the robot (see below).

Use MANUAL first to confirm motion and camera streams.

Why COLLECT may be greyed out

No collection config → COLLECT disabled (DEBUG, MANUAL, REPLAY still work).
Collection config present → COLLECT enabled.

Relaunch with a collection config; see below.

Workflow

#	Step	Action
1	TASK	Pick the task (list comes from `collection.tasks`). Saved with the episode; each task gets its own folder.
2	START	Recording begins. Status: `COLLECTING`; frame count climbs.
3	STOP	Episode is handed to the background saver. Status: `SAVING` until written.
4	QC	Replay and mark PASS/FAIL, with optional note.

What gets recorded. Both halves of the teleop: observed state (joint angles + EEF) and commanded action (joint angles + EEF). Driven by the config, not the tab.

While recording

START — recording begins immediately. Save and image encode run on a background worker; capture never blocks.
Backlog — if the host can't keep up, the queue grows and a warning is logged; recording continues.
Warm-up frames dropped — initial frames before the leader is tracking carry no action and are skipped.
STOP — episode is validated, packed, and queued for save; you can start the next one immediately. Empty episodes (0 frames) are discarded.
Cancel — discards everything captured so far.

Status line: COLLECTING, SAVING, QUEUE_FULL, or IDLE, plus frame count, queue depth, and ETA.

Recordings append. Relaunching into the same folder appends new episodes after existing ones. See Data collection for the on-disk layout.

Success and common faults

Symptom	Cause / action
Frame count stuck at 0	No data arriving. Verify drivers and that config topic names match the publishers.
Camera black / missing	Stream down or camera name mismatch. Confirm via MANUAL.
`QUEUE_FULL`	Saver behind. Wait for `IDLE`; if persistent, lower `fps` or raise `save_queue_max`.
Episode red on review	An auto-check fired — see Quality control.

Quality control

On STOP, replay the episode and mark PASS/FAIL. In addition, every episode runs through automatic checks; any failure marks it red and persists the reasons alongside the data.

Flag	Raised when	Action
`episode_too_short`	Fewer frames than `min_episode_frames`.	Re-record longer, or lower the threshold.
`non_monotonic_timestamp`	A frame's timestamp does not advance past the previous.	Re-record. If recurring, check the publisher clock.
`missing_camera` · `invalid_image_shape`	Configured camera absent, or image not H×W×3.	Verify stream and config name match.
`missing_configured_column` · `invalid_vector_dim`	Configured field missing or wrong length.	Verify topic publishing and arm count.
`non_finite_value`	NaN/Inf in a recorded value.	Auto-zeroed; re-record if frequent.
`frame_count_mismatch`	Camera video frame count differs from the data table.	Re-record; indicates dropped/stalled camera.

Bad fields are zero-filled, not dropped. Schema stays consistent; the flag is persisted with the episode.

For deeper review and adding instruction text per episode, use REPLAY.

Writing a collection config

A config becomes a collection config the moment it declares columns to record — that is what enables the tab. A collection config inherits a per-robot deploy file via _base_ and adds a collection section (see Configuration).

Settings

Setting	Purpose
`storage.log_dir`	Output root. Per-task subfolders. Empty → default folder named after the config.
`storage.fps`	Target frame rate.
`storage.save_queue_max`	Max episodes queued for save. Full → `QUEUE_FULL`.
`schema.robot_type`	Robot id, e.g. `agilex_AGILEX`.
`schema.min_episode_frames`	Threshold for `episode_too_short`.
`schema.arms`	Arms to record, one entry each.
`schema.cameras`	Cameras to record. At least one required.
`schema.columns`	The four required dataset fields: `qpos`, `eef`, `action_qpos`, `action_eef`.
`transport.ros1.groups`	Per-arm source topics for observed and commanded state. Use `ros2` for ROS 2.
`tasks`	Task strings offered in the TASK picker.

Copy the shipped collection config for your robot (e.g. configs/02_collection/dual_agilex_AGILEX.py) and edit the marked values. Lines marked ← change… must match your robot.

configs/02_collection/my_robot.py

_base_ = ['<path to your robot's deploy _base.py>']  # ← point at your robot's deploy config

collection = dict(
    storage=dict(
        log_dir='',           # ← where to save (empty = a default folder named after this config).
        fps=30,               # sane default; change only if your robot runs at another rate.
        save_queue_max=15,     # sane default.
    ),
    schema=dict(
        robot_type='<your robot type>',  # ← e.g. 'agilex_AGILEX'.
        min_episode_frames=10,  # sane default.
        arms=dict(left_arm='left', right_arm='right'),  # ← one entry per arm.
        cameras=dict(
            cam_high='observation.images.cam_high',  # ← list your cameras (name → dataset field).
        ),
        columns=dict(           # keep these four keys; values are standard dataset field names.
            qpos='observations.state.qpos',
            eef='observations.state.eef',
            action_qpos='action.qpos',
            action_eef='action.eef',
        ),
    ),
    transport=dict(
        ros1=dict(groups=dict(
            left_arm=dict(
                qpos_topic='<arm joint topic>',          # ← change to your robot's real topics.
                eef_topic='<arm gripper-pose topic>',
                action_qpos_topic='<commanded joint topic>',
                action_eef_topic='<commanded gripper-pose topic>',
            ),
            # ...repeat for each arm.
        )),
    ),
    tasks=['<your task instruction>'],  # ← e.g. 'pick up the apple'.
)

Save under configs/02_collection/ and launch. The loader enforces:

All four columns keys present — qpos, eef, action_qpos, action_eef.
cameras lists at least one camera.
arms lists at least one arm.

Field lengths are checked. Joint fields must match the robot's joint count; EEF fields are 8-wide per arm (x, y, z, orientation, gripper). Mismatches are flagged on review.

Launching

bash

eva --config configs/02_collection/my_robot.py

Open the console at http://localhost:8080 (override via --web-port, see CLI reference). Episodes are replayable through the dataset transport and reviewable in REPLAY.