Command-Line Reference

Run rpent --robot <name> --help for the full options of an environment. RPent supplies shared options; each environment adds task selection, resource paths, and service settings.

rpent --help
rpent --robot libero --help
rpent --robot robocasa --help
rpent --robot robotwin --help

Common Options

These options control the planner, memory, output directory, and Dashboard.

Option	Default	Meaning
--robot	—	Select a discovered robot or simulator.
--planner	api	api, claude_code, codex, or flash.
--model	—	api requires a provider prefix; SDK planners use their backend defaults.
--max-turns	100	Planner turn limit. For api, counts model requests across the conversation, including retries and follow-ups.
--max-tokens	8192	api only: token limit per model response.
--reasoning-effort	api: high; claude_code/codex: none	Reasoning effort for api, claude_code, and codex: none, low, medium, high, or xhigh. Disabling reasoning reduced the average runtime from approximately 13.2 to 7.9 minutes (about 40%) in our LIBERO Pro Long evaluations. Higher effort may improve task success rate. Supported levels ultimately depend on the selected model.
--planner-timeout-s	—	Defaults to CODEX_TIMEOUT_S (Codex only), then CELL_TIMEOUT_S, then 1200 seconds; terminal interactive API/Claude sessions are exempt.
--base-url	—	api only: override the model endpoint. SDK planners use their own environment variables.
--no-images	false	api only: omit images from model requests.
--claude-code-max-budget-usd	—	Claude Code dollar budget; defaults to MAX_BUDGET_USD or 10.
--output-dir	—	Artifact directory; the environment generates the default name.
--memory-profile	—	Evaluation normally defaults to hf; exploration and WAM evaluation use local.
--memory-dir	—	Local memory directory; use with the local profile or exploration.
--memory-repo	—	HF dataset ID or URL with optional @revision; evaluation only.
--explore	false	Enable exploration on supported environments.
--tui, -t	false	Enter instructions in the terminal; see Terminal Interaction (TUI).
--dashboard	false	Start a Dashboard session.
--dashboard-host	127.0.0.1	Dashboard bind address.
--dashboard-port	0	0 selects an available port.
--dashboard-language	en	en or zh-cn.
--verbose	false	Enable DEBUG logging.

Environment Options

LIBERO selects tasks with --suite, --task, --seed, and --libero-type; the variant defaults to LIBERO_TYPE or pro. RoboCasa365 and RoboTwin use --task-name. Franka uses --task-id and --robot-config. See LIBERO, RoboCasa365, RoboTwin, Single-Arm Franka, and Dual-Arm Franka for details.

--env-endpoint, --am-endpoint, and LIBERO’s --sam3-endpoint connect to existing environment, action-model (VLA or WAM), and SAM3 services respectively. RPent RPC addresses use [protocol://]host:port, defaulting to http. LingBot uses native WebSocket; see RoboTwin. See Remote Services and Parallel Runs for GPU assignment and remote services.

--am-backend selects one action model supported by the platform. LIBERO defaults to pi05 and RoboTwin to lingbot. VLA and WAM share --am-endpoint for existing workers, and --am-checkpoint, --am-python, --am-root and --am-cuda-device for local deployment. Explicit local deployment options are rejected with an endpoint. --am-checkpoint supplies local weights or a service model identity as required by the backend. --am-python selects the provisioned model interpreter; VLA defaults to the current interpreter. --am-root selects the working directory, defaulting to the RPent checkout; Cosmos Policy requires its source directory.

WAM uses the local memory profile with memory file tools and a backend-specific directory; --memory-dir overrides it. WAM currently rejects HF memory, Flash and exploration. See Add an Action Model for model launch options and supported pairs.

--predict-future enables future-state/value outputs during local deployment. Currently only Cosmos Policy supports this option; other backends reject it. Configure it when starting the worker if connecting with --am-endpoint. RPent does not yet use these outputs for planning.

--am-cuda-device selects the local model GPU independently of the simulator. Local model deployment is supported by LIBERO, RoboTwin, RoboCasa, RoboDojo and dual-Franka. Franka and YAM connect to externally deployed VLA services; select their GPU when deploying the service. LIBERO, RoboTwin, RoboCasa and RoboDojo use --env-cuda-device for the environment. LIBERO, RoboDojo and dual-Franka use --sam3-cuda-device for local SAM3. These services may share a GPU when memory permits. External endpoints use the GPU configuration set when their services were started. The original --cuda-device option remains a shared default for local services. Per-service options take precedence, independent of argument order; without either option, services inherit CUDA_VISIBLE_DEVICES. For example, --cuda-device 0 --am-cuda-device 1 starts the environment and SAM3 on GPU 0 and the WAM worker on GPU 1. Standalone service commands retain their own --cuda-device where supported. rpent --robot libero --am-backend fast-wam --help groups connection, local deployment and selected-model options. Model-specific options retain their model prefix; options for another model are rejected. The action-model guide separates one-time worker deployment from repeated task commands and includes examples for both external services and workers started by RPent.

The following defaults apply to LIBERO; check each other environment’s --help for its defaults.

LIBERO option	Default	Meaning
--seed	0	Environment random seed.
--memory-version	auto	HF corpus: auto, GPT_5.5_xhigh, or GPT_6_astra_low. See Memory and Exploration; explicit versions cannot be combined with local memory or exploration.
--max-episode-steps	10000	Environment step limit, separate from planner turns (--max-turns).

Action-Model Compatibility

Use the current options in the left column for new commands. The right column lists hidden aliases retained for existing commands; they do not appear in --help and write the same configuration fields as their current names.

Common deployment options use --am-*. Model-specific options use --pi05-*, --cosmos-*, --cosmos3-*, --fast-wam-* or --lingbot-*. Here fast-wam is the model name, distinct from the old generic --wam-* prefix.

Model option names

Current option	Compatibility alias (hidden)
--am-backend	--vla-backend, --wam-backend
--am-backend pi05	--policy-backend rlinf (RoboDojo)
--am-endpoint	--vla-endpoint, --wam-endpoint
--am-checkpoint	--vla-model-path, --wam-checkpoint
--am-python	--pi05-python, --wam-python
--am-root	--wam-root
--am-cuda-device	--vla-cuda-device, --wam-cuda-device
--pi05-repo-id	--vla-repo-id
--cosmos-text-embeddings	--wam-text-embeddings
--cosmos-cached-instructions-only	--wam-cached-instructions-only
--fast-wam-config	--wam-config
--fast-wam-dataset-stats	--wam-dataset-stats

Model-Service Environment Variables

Set the credentials and endpoint for your chosen planner in the launch shell.

Planner	API key	Endpoint
api / Anthropic; claude_code	ANTHROPIC_API_KEY	ANTHROPIC_BASE_URL
api / OpenAI	OPENAI_API_KEY	OPENAI_BASE_URL
codex	CODEX_API_KEY	CODEX_BASE_URL

Codex can also reuse existing Codex authentication. See Planner Configuration for model selection, endpoint precedence, and local server examples. Use rpent-check-llm for connectivity checks and rpent-memory for the operations in Memory and Exploration.

Each environment reads its own resource variables, including LIBERO’s PI05_CHECKPOINT_PATH and SAM3_CHECKPOINT_PATH, RoboCasa365’s ROBOCASA_ASSETS_PATH, and RoboTwin’s ROBOTWIN_ASSETS_PATH and LINGBOT_MODEL_PATH.

Output Files

--output-dir selects the run directory. Without it, LIBERO writes to logs/<timestamp>_<suite>_t<task>_s<seed>/ under the repository; other environments describe their directory names on their own pages.

The table uses LIBERO as an example. Files depend on the environment, recording, and run outcome.

File or path	Purpose
transcript_*.json	Planner conversation, tool calls, and completion information.
episode.mp4	Episode video.
run.log	RPent process log.
env_server.log / vla_server.log / wam_server.log / sam3_server.log	Logs from environment and model servers started by this run.
*_recipe.jsonl	Action sequence exported from a successful task; failed runs may not produce this file.
states.json	Internal EnvState manifest. Inspect states through the Dashboard or view_env_state; callers should not parse this file directly.
agentview_depth.npz/00.npz	Example step artifact: the directory uses the logical artifact name and contains numbered steps such as 00.npz and 01.npz.

Run-level files live at the output root; step observations live in their own subdirectories. Use the native success criterion documented for each environment; a recipe, video, or planner summary does not replace it.