Skip to content

World Model Evaluation Service ​

Updated: 2026-11-11

This guide explains how to call the world model evaluation service and run tests with the Soma Policy client. The service endpoint is http://129.204.20.25:18085. Request IP access from the service contact before connecting. The current API uses HTTP and does not require an SSH tunnel or API token.

A rollout follows this loop: choose a scene → get images and state → predict actions with your Policy → receive new images and state from the world model → continue.

State values

Images are transferred as PNG Base64 strings. The SDK can decode them into RGB images. Subsequent state values are currently estimated from the last executed action; they are not measured from the generated image. Check the meaning of gripper values carefully when connecting a different Policy.

Interface Contract ​

The world model provides the initial images and state for a scene, then returns new images, state, and video after actions are executed. Images include the main view, left wrist, and right wrist cameras.

The Policy caller provides a scene ID and actions predicted from the current images and state. Initial images and state are already hosted by the service, so callers do not need to prepare a dataset.

Choose a scene ​

There are currently 20 scenes, numbered from 0, such as block_sorting_0 through block_sorting_19. Each ID has its own initial images and state. Change the ID to switch scenes.

Action format ​

Each request contains 24 rows of 14 values, ordered as follows:

text
6 left-arm joint angles + left gripper + 6 right-arm joint angles + right gripper
  • Joint angles are in radians and represent absolute target angles.
  • Gripper commands range from 0 to 1.
  • The action frequency is 20 Hz, so each row represents 0.05 seconds.
  • Send a 24-row plan each time. You may execute only the first 16 rows, advancing 0.8 seconds before predicting again.

The 20 Hz value describes the spacing of actions in the sequence; it does not require 20 network requests per second. If your Policy outputs hand positions, relative actions, or another format, convert them to the joint-action format first. The Soma client includes this conversion.

API Endpoints ​

Append these paths to the service endpoint. For example, the readiness check is http://129.204.20.25:18085/v1/ready.

MethodPathPurpose
GET/v1/readyCheck whether the service can accept requests
GET/v1/tasksList available scenes
GET/v1/tasks/block_sorting_7Get information about a scene
POST/v1/sessionsSelect a scene and start a rollout; returns the initial images and state
POST/v1/sessions/{session_id}/stepsSubmit Policy actions
GET/v1/jobs/{job_id}Check generation status and retrieve new images and state
GET/v1/artifacts/{artifact_id}Download the round's video
DELETE/v1/sessions/{session_id}End the rollout

Start a rollout by sending:

json
{"task_id":"block_sorting_7"}

session_id identifies the full rollout; job_id identifies one generation round. The service returns both IDs. After an action is submitted, the service first returns a job ID. Query the job endpoint after generation completes to retrieve its result. The SDK handles polling for you.

Python SDK ​

Contact the service owner for the SDK 0.2.0 delivery package and wheel. Install it with:

bash
python -m pip install ./lumos_wm_sdk-0.2.0-py3-none-any.whl

The example below shows the call sequence. actions20 is a 24×14 action array generated and converted by your Policy. This is an integration example, not a complete Policy program.

python
from lumos_wm import WorldModelClient

client = WorldModelClient("http://129.204.20.25:18085")

# 1. Select scene 7 and get its initial images and state
session = client.reset_task("block_sorting_7")
observation = session["observation"]

# 2. Run your Policy with observation to generate actions20

# 3. Submit actions and wait for new images and state
result = client.step(
    session["session_id"],
    actions20,
    expected_step_index=0,
    execute_steps=16,
)
observation = result["observation"]

# 4. Continue predicting from new observations, then close the session
client.close(session["session_id"])

You can use your own Policy; Soma is not required. Contact the service owner for the full parameter reference, error handling, and more examples.

Run a Local Soma Test ​

The commands below are for the current test environment. The world model runs persistently on its server and does not need to be restarted for each test.

1. Start the local Policy ​

Run this in terminal A and leave it running. If the same service is already running, reuse it:

bash
cd /home/ubuntu/work/nexworld_eval
/home/ubuntu/work/LumosVLA/newdeploy/soma configs/soma_block_wm.toml

This loads the block-task checkpoint so the Policy can predict actions from images and state.

2. Check the connection ​

Run this in terminal B:

bash
cd /home/ubuntu/work/nexworld_eval
curl --noproxy '*' -fsS http://129.204.20.25:18085/v1/ready

accepting_requests=true means the service can accept requests.

3. Run a two-round test ​

Continue in terminal B:

bash
# Connect directly to the world model service instead of forwarding through a local proxy
export NO_PROXY="${NO_PROXY:+$NO_PROXY,}129.204.20.25,127.0.0.1,localhost"
export no_proxy="$NO_PROXY"

TEST_OUT="outputs/block_sorting_0_public_$(date +%Y%m%d_%H%M%S)"
/home/ubuntu/miniconda3/bin/python scripts/soma_wm_client.py \
  --config configs/soma_wm_client_public.toml \
  --task-id block_sorting_0 \
  --max-windows 2 \
  --output "$TEST_OUT"

This runs two rounds of Policy prediction and world model generation, advancing 1.6 seconds of simulated time. Requests go directly to the public endpoint; no SSH tunnel is needed. To change the scene, replace block_sorting_0. To run the full reference duration, remove --max-windows 2 and use a new output directory. Scene 0 advances 50.7 seconds by default; other scenes use their respective data durations.

Test Results ​

Results are saved to the local directory specified by --output:

  • rollout.mp4: stitched video.
  • summary.json: number of rounds, simulated duration, and completion state.
  • One folder per round: action records, returned images, and round video.

After a successful two-round test, the terminal reports 1.6 seconds advanced and Rollout complete. Actual generation takes longer than the video duration in most cases. The video includes an initial frame, so the two-round video is about 1.65 seconds for 1.6 seconds of simulated progress.

Notes ​

  • The current API uses HTTP. Request IP access first; SSH is not required to call the API.
  • A single model processes jobs in sequence, so concurrently submitted jobs wait in a queue.
  • Subsequent state values approximate action targets and may not exactly match the image.
  • The service does not automatically determine whether the block task succeeded. A completed rollout does not mean the task succeeded.
  • Existing sessions cannot be resumed after a service restart; start a new rollout.

Contact the service owner for additional operations, recovery steps, field definitions, response formats, deployment, and operations guidance.

NexCore is evolving rapidly; the latest version prevails.NexCore Docs · Embodied AI Skill Learning Platform by Lumos Robotics