# AI use cases for Robo Harness (BinceQu/RoboHarness)

Repo: [BinceQu/RoboHarness](https://github.com/BinceQu/RoboHarness) | Use cases: 3 | Page: https://dedupedai.com/use-cases/bincequ-roboharness/

RoboHarness is a Python harness that lets LLM agents control simulated robot manipulation tasks in BEHAVIOR through a visual-geometric control panel, using tracked image keypoints and depth back-projection.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- Reproduce a BEHAVIOR task mean Q-score: https://dedupedai.com/use-cases/bincequ-roboharness/reproduce-behavior-task-mean/
- Watch one agent rollout in a browser: https://dedupedai.com/use-cases/bincequ-roboharness/watch-single-instance-rollout/
- Point the harness at your own model endpoint: https://dedupedai.com/use-cases/bincequ-roboharness/run-harness-on-own-model-endpoint/

## How to reproduce a BEHAVIOR task mean Q-score with Robo Harness

Repo: [BinceQu/RoboHarness](https://github.com/BinceQu/RoboHarness) | Difficulty: Advanced | Time: 8 hours | Basis: From the README

Run the five fixed BEHAVIOR instances for one task and compare the mean with the reference value.

**Who it is for:** A robotics researcher checking a published embodied-agent result on their own GPU machine before building on it.

**The problem:** Published numbers for LLM-driven manipulation are hard to trust without rerunning the same instances under the same protocol.

**The outcome:** A validation report in .local/reports/ that shows whether the task mean matches the reference within the stated tolerance.

### Steps

1. `git clone --recurse-submodules https://github.com/BinceQu/RoboHarness.git`
2. `./scripts/setup.sh`
3. `cp configs/example.json configs/local.json`
4. `./run.sh --list`
5. `./scripts/reproduce_task.sh task01 --gpu 0 --dry-run`
6. `./scripts/reproduce_task.sh task01 --gpu 0 --run-dir runs/task01-001`
7. `python3 scripts/report_validation.py runs/task01-001 --output .local/reports/task01-001 --check-live --require-match` (inferred, check the README)

### Prompt to paste into your agent

```text
In the RoboHarness repo, run the task <task id from ./run.sh --list> with --dry-run first, then the full five-instance run into a new run directory, then generate the validation report. Report the task mean from runs/<run>/summary.json and whether --require-match passed. Verify that all five instance IDs are 301, 304, 306, 308 and 310 before you start.
```

### Caveat

The full run needs licensed BEHAVIOR data and a separately served model, and the README gives 48 GB VRAM per GPU only as the tested capacity, not a minimum.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/bincequ-roboharness/reproduce-behavior-task-mean/

## How to watch one agent rollout in a browser with Robo Harness

Repo: [BinceQu/RoboHarness](https://github.com/BinceQu/RoboHarness) | Difficulty: Advanced | Time: 4 hours | Basis: From the README

Run a single BEHAVIOR instance and open the live interface through an SSH tunnel.

**Who it is for:** A researcher debugging why an LLM agent fails a household task, working on a GPU server they reach over SSH.

**The problem:** Full five-instance runs hide the step-by-step behavior of the agent, which is needed to diagnose failures.

**The outcome:** A separate run directory with the prompt, agent output, error log and trajectory for instance 301, viewed live in the browser.

### Steps

1. `ssh -N -L 15071:127.0.0.1:15071 YOUR_USER@YOUR_EVALUATION_HOST`
2. `./scripts/reproduce_task.sh task01 --gpu 0 --run-dir runs/task01-301 --instances 301` (inferred, check the README)
3. Open http://127.0.0.1:15071 in your own browser to watch the rollout.
4. Read `runs/task01-301/instance_301/` for the prompt, agent output, error log and trajectory.

### Prompt to paste into your agent

```text
In the RoboHarness repo, start a single-instance trial of task <task id> with --instances 301 in a new run directory that does not already exist. Report the path of instance_301/ and the last lines of its error log. Verify that the run directory is new before launching.
```

### Caveat

The README says a single-instance result is only a trial, because the full task mean requires all five instances.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/bincequ-roboharness/watch-single-instance-rollout/

## How to point the harness at your own model endpoint with Robo Harness

Repo: [BinceQu/RoboHarness](https://github.com/BinceQu/RoboHarness) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README

Connect the Claude Code harness to a self-hosted Anthropic-compatible model server.

**Who it is for:** An ML engineer who serves an open model on their own hardware and wants to test it as the control agent for robot tasks.

**The problem:** The harness expects a specific endpoint type, and misconfigured URLs or model IDs fail before any robot motion starts.

**The outcome:** A verified connection to the model server, a saved session configuration, and a dry-run plan that shows the correct model URL and model ID.

### Steps

1. `export ROBOHARNESS_MODEL_URL='http://YOUR_MODEL_HOST:31000'`
2. `export ROBOHARNESS_MODEL='Qwen3.8-Flash-Next-FP8'`
3. `export ANTHROPIC_API_KEY='YOUR_MODEL_SERVER_KEY'`
4. `export ANTHROPIC_AUTH_TOKEN="$ANTHROPIC_API_KEY"`
5. Send the small test request in the README and expect a Messages response with "type": "message".
6. `./scripts/reproduce_task.sh task01 --gpu 0 --dry-run`

### Prompt to paste into your agent

```text
In the RoboHarness repo, set the model URL and model ID from the environment variables for <model server host>, send the small Messages test request from the README, and then run the task01 dry run. Report the model_url, model, gpu and port values from the dry-run output. Verify that the URL has no /v1 suffix before you run the test.
```

### Caveat

The README says dry-run does not contact the model or check the simulator installation, so a successful dry run does not prove the endpoint works.

Risk, credentials: The model server key is exported into the shell environment, so keep it out of shared logs and Git checkouts.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/bincequ-roboharness/run-harness-on-own-model-endpoint/