Task Catalog

Use the task catalog to find a logical task, select a runnable deployment, and inspect the evidence available for it.

Discover and inspect tasks

embodichain list-task --category manipulation
embodichain show-task embodichain_tasks:repeated_pick_place

show-task reports the task description, named deployments, robot embodiment, physics backend, supported uses, configuration paths and supplied validation records. Use the qualified package:task_key when different packages share a task key. Each displayed launch command selects a concrete configuration.

The catalog distinguishes three artifacts:

Artifact	Responsibility	Example
Task catalog	Human-facing identity and named deployment references	catalog.yaml
Runnable deployment	Gym ID and selections of environment, embodiment and optional program	task.ur5.yaml

Catalog metadata does not register a Gym ID. Existing tasks without metadata remain discoverable. Capabilities describe available execution routes; they are not evidence that a deployment passed a physical evaluation. In particular, RL support is associated with the deployment referenced by a trainer configuration, not every deployment in a directory containing an agents folder.

Export a local task gallery

From a source checkout, export a static HTML gallery without importing simulator registries:

embodichain list-task \
  --config-root embodichain_tasks=embodichain_tasks/configs/tasks \
  --category manipulation \
  --export-html task-gallery.html

Open task-gallery.html in a browser. --config-root points to a configs/tasks directory; repeat it for additional packages. This static mode does not inspect runtime registrations, so additional registered capabilities may be available. Omit --config-root to use installed task discovery.

The generated gallery links local task resources where available. Keep those resources in place when using the exported file; it is not a portable archive of all configurations. Missing previews and validation results are shown as unavailable.

The repeated-pick-place catalog is authored alongside its deployments:

task_key: repeated_pick_place
title: Repeated Pick and Place
summary: Move a cube between two locations using a repeated Task Program.
tags: [pick_place, single_arm, task_program]
default_deployment: franka
deployments:
  franka:
    config: task.franka.yaml
  ur5:
    config: task.ur5.yaml
  ur5_objective:
    config: task.ur5.objective.yaml

default_deployment names a declared deployment. Each config references a task-relative JSON, YAML or YML runnable configuration; optional validation references a task-relative JSON report. Select a different deployment to change physics backends: launcher --physics does not override the backend owned by the physical environment component.

Measure an independent physical objective

The ur5_objective deployment attaches an optional ordered stable-region objective to the repeated pick-place task. It observes the cube reaching the configured regions in order and reports physical progress separately from Task Program completion, segment acceptance and dataset persistence.

python -m embodichain.lab.scripts.evaluate_task_objective \
  --gym_config embodichain_tasks/configs/tasks/manipulation/repeated_pick_place/task.ur5.objective.yaml \
  --output-dir objective-run --seed 0 --num_envs 1 --headless --device cuda

The report records the resolved deployment, component hashes, effective seed, actual initial pose, expert outcome, dynamic replay outcome and physical objective snapshot. Add --initial-position-jitter 0.005 for a bounded initial pose variation. This first runner supports one environment and dynamic replay through the recorded horizon; a compatible DexSim build is required for native physical qualification.

For launch and recording options, see Running Environments with run-env.