Quick Start#

Before you begin, follow Installation to install RPent and download the LIBERO-PRO simulator assets. The steps below use LIBERO-PRO with the claude_code planner to demonstrate a complete run.

1. Configure keys and checkpoints#

Export your Anthropic key, then download and configure the VLA and SAM3 checkpoints:

# Anthropic key; no need to export the base url if you use the
# official endpoint.
export ANTHROPIC_BASE_URL=https://xxx
export ANTHROPIC_API_KEY=sk-xxx

# VLA checkpoint — download from
# https://huggingface.co/RLinf/RLinf-Pi05-LIBERO-130-fullshot-SFT
pip install "huggingface_hub>=0.34,<1.0"

hf download RLinf/RLinf-Pi05-LIBERO-130-fullshot-SFT \
  --exclude optimizer.pt \
  --local-dir ./checkpoints/RLinf-Pi05-LIBERO-130-fullshot-SFT

export PI05_CHECKPOINT_PATH=$PWD/checkpoints/RLinf-Pi05-LIBERO-130-fullshot-SFT

# SAM 3.0 checkpoint — download from
# https://modelscope.cn/models/facebook/sam3
pip install -U modelscope

modelscope download facebook/sam3 \
  --local-dir ./checkpoints/sam3

export SAM3_CHECKPOINT_PATH=$PWD/checkpoints/sam3/sam3.pt

2. Run one LIBERO task#

Run a single LIBERO PRO task (libero_object_swap, task 2, seed 0) using the claude_code planner:

rpent --robot libero --suite libero_object_swap --task 2 --seed 0 \
  --planner claude_code --model claude-opus-4-8

To switch to another planner, such as codex or api, see Agentic Planner.

3. Monitor the run in the Dashboard#

Add --dashboard to start a local Dashboard and print its URL in the terminal:

rpent --robot libero --dashboard --dashboard-language zh-cn \
  --planner claude_code --model claude-opus-4-8

Session configuration comes from the CLI and the URL opens directly in the live monitor. Once the services are ready, enter /rpent-task libero_object_swap 2 0 in the page to start a task. The Dashboard streams agent reasoning, camera views, and the action timeline; submit another task after the current one finishes. Use --dashboard-language zh-cn for the Chinese UI.

Key CLI options#

The table lists the main CLI options. Run rpent --help for other general options. See the LIBERO guide for detailed robot configuration.

Main

Flag	Default	Description
`--robot`	— (required)	Robot backend. Currently `libero`.
`--suite`	— (required)	Task suite, e.g. `libero_object_task`, `libero_spatial_swap`
`--task`	— (required)	Task id within the suite
`--seed`	`0`	Random seed
`--libero-type`	`LIBERO_TYPE` or `pro`	LIBERO variant: `standard` \| `pro` \| `plus`

Planner

Flag	Default	Description
`--planner`	`api`	`api` \| `claude_code` \| `codex`
`--model`	—	Model id; for `api`, prefix the provider (`anthropic:…`, `openai:…`, `openai-chat:…`)
`--max-turns`	`100`	Max agent turns
`--max-tokens`	`8192`	Max tokens per LLM reply
`--reasoning-effort`	`none`	Reasoning effort for `api`, `claude_code`, and `codex`: `none` \| `low` \| `medium` \| `high` \| `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.
`--no-images`	off	Text-only mode: never send image bytes (for models that reject image input)

Environment

Flag	Default	Description
`--max-episode-steps`	`10000`	Max env steps
`--cuda-device`	inherited	GPU device exposed to the env / VLA / SAM3 servers
`--env-endpoint`	— (spawn)	`[protocol://]host:port` of an existing env_server (`protocol=http\|socket`, default `http`). If unset, one is spawned locally.
`--vla-endpoint`	— (spawn)	`[protocol://]host:port` of an existing vla_server (same rules). If unset, one is spawned locally.
`--sam3-endpoint`	— (spawn)	`[protocol://]host:port` of an existing RPent SAM3 service (same rules). If unset, one is spawned locally.

Dashboard

Flag	Default	Description
`--dashboard`	off	Start a local Dashboard
`--dashboard-language`	`en`	Dashboard UI language: `en` \| `zh-cn`

What you should see#

A successful run:

Shows startup messages for env_server, vla_server, and sam3_server in the terminal.
Prints per-turn agent output and tool calls in the terminal, followed by the elapsed time, token usage, and path to the run record.
With the Dashboard enabled, also streams agent output, camera views, the action timeline, and clip replays to the Dashboard.
By default, artifacts are saved under logs/<timestamp>_<suite>_t<task>_s<seed>/. They include transcript_*.json (run record), states.json (the EnvState manifest), *_recipe.jsonl (action sequence), and episode.mp4 (episode video). Each step artifact has a directory named after its logical artifact name; zero-padded step files live inside it, for example agentview_depth.npz/00.npz and agentview_depth.npz/01.npz. Run-level artifacts remain at the output root.

Inspect the final state through the Dashboard or view_env_state(step=-1). Its top-level terminated value is the benchmark outcome. states.json is internal EnvState storage and should not be parsed by callers. You can also open episode.mp4 to review the run. If something goes wrong, inspect the four log files described at the bottom of Installation.