Configuration
Robo Use reads its settings from environment variables; the ones you set are listed here, generated from robouse.config. Paths default to ~/.cache/robouse (the cache) and ~/.config/robouse (private settings such as remote worker endpoints, which the local runner's sandbox keeps from agents).
| Variable | Meaning |
|---|---|
ROBOUSE_CACHE_DIR | cache for simulator virtualenvs, downloaded assets and harness homes (default ~/.cache/robouse) |
ROBOUSE_CONFIG_DIR | private settings such as remote worker endpoints (default ~/.config/robouse); agents never see it |
ROBOUSE_MENAGERIE_ASSETS | MuJoCo Menagerie robot models (default: the checkout's assets/menagerie, else <cache>/menagerie) |
ROBOUSE_PRICING | JSON price table for trial cost and the spend cap, used instead of the packaged pricing.json (not added to it) |
ROBOUSE_RECORD_SIZE | episode video size WIDTHxHEIGHT (default: the backend's native size) |
ROBOUSE_AGENT_IMAGES | 0: the model gets no image input in this run, whatever it takes (its gateway replaces every image part; recorded as image_input: false). Default: from the model registry |
ROBOUSE_SPEND_CAP_USD | the model spend at which the model gateway stops a trial, in USD (default 5; 0: no cap; recorded as a budget stop). A model without a price, such as glm/glm-5.3, runs with 0 or with its price in ROBOUSE_PRICING |
ROBOUSE_SANDBOX | 0 disables the sandbox around agent processes (macOS Seatbelt, Linux bubblewrap; default 1) |
ROBOUSE_SANDBOX_DENY | extra colon-separated paths agents may neither read nor write |
ROBOUSE_DEXJOCO_PYTHON | Python of the DexJoCo virtualenv (default <cache>/dexjoco-venv/bin/python) |
ROBOUSE_DMCONTROL_PYTHON | Python of the dm_control virtualenv (default <cache>/dmcontrol-venv/bin/python) |
ROBOUSE_HUMANOIDBENCH_PYTHON | Python of the HumanoidBench virtualenv (default <cache>/humanoidbench-venv/bin/python) |
ROBOUSE_MANISKILL_PYTHON | Python of the ManiSkill virtualenv (default <cache>/maniskill-venv/bin/python) |
ROBOUSE_METADRIVE_PYTHON | Python of the MetaDrive virtualenv (default <cache>/metadrive-venv/bin/python) |
ROBOUSE_MYOSUITE_PYTHON | Python of the MyoSuite virtualenv (default <cache>/myosuite-venv/bin/python) |
ROBOUSE_ROBOCASA_PYTHON | Python of the RoboCasa virtualenv (default <cache>/robocasa-venv/bin/python) |
ROBOUSE_RLEBENCH_PYTHON | Python of the RLE-Bench virtualenv (default <cache>/rlebench-venv/bin/python) |
ROBOUSE_CAPX_PYTHON | Python of the CaP-X virtualenv (default <cache>/capx-venv/bin/python) |
ROBOUSE_CAPX_LIBERO_PYTHON | Python of the CaP-X LIBERO virtualenv (default <cache>/capx-libero-venv/bin/python) |
ROBOUSE_YAM_PYTHON | Python of the dual-YAM virtualenv (MuJoCo 3.3.5+ and Dream Labs' scene; default <cache>/yam-venv/bin/python) |
ROBOUSE_VLA_URL | policy server for the vla harness |
ROBOUSE_VLA_TOKEN | bearer token for the policy server |
ROBOUSE_VLA_POLICY_CONFIG | JSON file with overrides for the vla preset |
ROBOUSE_DREAM_PYTHON | interpreter that runs the dream client, with Dream's client library installed (default <cache>/dream-venv/bin/python) |
ROBOUSE_DREAM_ACTIONS | rows of each Dream-1 action chunk the dream client executes before asking again (1-48, default 48) |
ROBOUSE_MOLMOACT2_URL | MolmoAct2 policy server for the molmoact2 harness |
ROBOUSE_MOLMOACT2_TOKEN | bearer token for the MolmoAct2 server |
ROBOUSE_DIMCODE | the dimcode executable for the dimcode harness (default: dimcode on PATH) |
ROBOUSE_DIMCODE_NODE | Node 24 or 26 that runs dimcode (default: node on PATH) |
ROBOUSE_SHOW_HARNESS_URL | OpenAI-compatible server (base URL ending in /v1) of a self-hosted model for the show-harness harness, e.g. vLLM serving a Show-Harness adapter |
ROBOUSE_SHOW_HARNESS_TOKEN | token for that server: a bearer token, or sent under the overrides' token_header |
ROBOUSE_SHOW_HARNESS_CONFIG | JSON file of show-harness overrides (mode zs or ft, step sizes, plugins) |