Quick start#

This page ports the README.md Quick Start into the documentation. It assumes you have already followed Installation (RLinf and RPent cloned side by side, the virtualenv activated, RPent’s extras synced).

1. Configure keys and checkpoints#

Export the API keys for the LLM provider(s) you want to use as the reasoning brain, plus the path to the VLA checkpoint:

# LLM API keys (used by the `api` cerebrum via pydantic-ai)
export ANTHROPIC_BASE_URL=https://xxx
export ANTHROPIC_API_KEY=sk-xxx
export OPENAI_BASE_URL=https://xxx
export OPENAI_API_KEY=sk-xxx

# VLA checkpoint — download from
# https://huggingface.co/datasets/RLinf/rlinf-pi05-libero-130-fullshot-sft
export PI05_CHECKPOINT_PATH=/path/to/rlinf-pi05-libero-130-fullshot-sft
export LIBERO_TYPE=pro
export CUDA_VISIBLE_DEVICES=0

You only need to set the keys for the providers you actually target. For example, if you only run --cerebrum claude_code, you can skip OPENAI_*.

2. Run one LIBERO task#

Run a single LIBERO PRO task (libero_object_swap, task 2, seed 0) using the api cerebrum against an Anthropic model with an 8192-token cap:

python rpent/cli/main.py --suite libero_object_swap --task 2 --seed 0 \
  --cerebrum api --model anthropic:claude-opus-4-8 --max-tokens 8192

Model id conventions. --model accepts a provider-prefixed id for the api cerebrum, and a bare id for the claude_code / codex cerebrums:

  • OpenAI-compatible chat endpoints — --model openai-chat:glm-5.2

  • OpenAI responses endpoints — --model openai:gpt-5.5

  • claude_code / codex — no provider prefix, e.g. --model claude-opus-4-8

See Agentic planner for the full brain-swapping guide.

3. Watch it run in the dashboard#

Add --dashboard to open a browser monitor for the run. It boots a launcher screen where you pick the config, then streams the agent’s reasoning, live camera and Pi0 views, an action timeline, and clip replays. Use --dashboard-language zh-cn for the Chinese UI.

python rpent/cli/main.py --dashboard --dashboard-language zh-cn \
  --suite libero_goal_task --task 1 --seed 0 --cerebrum claude_code

4. RoboCasa#

RoboCasa uses a separate entrypoint and one-time setup:

bash scripts/setup_robocasa.sh                                # one-time
bash scripts/run_robocasa.sh PickPlaceCounterToCabinet 0 0    # <task> <gpu> <seed>

See docs/SETUP_ROBOCASA.zh.md for the full RoboCasa365 + RLDX-1 walkthrough, and RoboCasa for what the RoboCasa toolkit exposes to the agent.

Key CLI options#

The most common flags of rpent/cli/main.py at a glance:

Flag

Default

Description

--suite

required

Task suite, e.g. libero_object_task, libero_spatial_swap

--task

required

Task id within the suite

--seed

0

Random seed

--cerebrum

api

Reasoning brain: 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

--max-episode-steps

10000

Max env steps

--libero-type

LIBERO_TYPE or pro

LIBERO variant: standard | pro | plus

--cuda-device

inherited

GPU device(s) exposed to the env / vla servers

--dashboard

off

Start the local dashboard for this run

--dashboard-language

en

Dashboard UI language: en | zh-cn

--vla-endpoint

—

Reuse an already-running vla_server instead of spawning one

--no-driver

off

Attach to an existing env_server / vla_server

What you should see#

A successful run:

  1. Prints env server ready at 127.0.0.1:<port> once the driver process is up.

  2. Prints per-turn agent reasoning (or streams it to the dashboard).

  3. Ends when the LLM calls finish(success=True), or hits --max-turns / --max-episode-steps.

  4. Writes <output_dir>/transcript_*.json with the full turn-by-turn record and <output_dir>/episode.mp4 with the rendered rollout.

If something goes wrong, inspect the three log files described at the bottom of Installation.