Advanced Deployment

RPent normally starts and stops the LIBERO environment, VLA, and SAM3 services with each run. Keep that default for a single-machine setup. Use external endpoints when the services run on different hosts or when the VLA and SAM3 models should be reused across runs.

The endpoint options support the following transports:

Service	RPent option	Endpoint format
LIBERO environment	--env-endpoint	HTTP or socket RPC, [protocol://]HOST:PORT
Pi0.5 VLA	--vla-endpoint	HTTP or socket RPC, [protocol://]HOST:PORT
SAM3	--sam3-endpoint	HTTP or socket RPC, [protocol://]HOST:PORT

LIBERO environment server

An environment server is tied to one suite, task, seed, and episode-step limit. These values must exactly match the RPent client command. On the environment host, run:

export LIBERO_TYPE=pro
export CUDA_VISIBLE_DEVICES=0
python -m robots.libero.env_server \
  --suite libero_object_swap --task 2 --seed 0 \
  --max-episode-steps 10000 \
  --transport http --host 0.0.0.0 --port ENV_PORT

The server is task-specific. Stop it and start a new one before changing any of the matching arguments.

Pi0.5 VLA server

On the VLA host, set the checkpoint path and start the HTTP service:

export PI05_CHECKPOINT_PATH=/path/to/rlinf-pi05-libero-130-fullshot-sft
export CUDA_VISIBLE_DEVICES=0
python -m robots.libero.vla_server \
  --transport http --host 0.0.0.0 --port VLA_PORT

The VLA server loads the model once and can be shared by multiple RPent runs.

SAM3 server

On the SAM3 host, set the local checkpoint path and start the HTTP service:

export SAM3_CHECKPOINT_PATH=/path/to/sam3/sam3.pt
export CUDA_VISIBLE_DEVICES=0
python -m robots.libero.sam3_server \
  --transport http --host 0.0.0.0 --port SAM3_PORT

The SAM3 server loads the model once and can be shared by multiple RPent runs.

Connect RPent

On the machine running RPent, use the three endpoint options to connect to the services above. The suite, task, seed, and maximum episode steps must match the environment server command:

rpent \
  --env libero \
  --suite libero_object_swap --task 2 --seed 0 \
  --libero-type pro --max-episode-steps 10000 \
  --env-endpoint http://ENV_HOST:ENV_PORT \
  --vla-endpoint http://VLA_HOST:VLA_PORT \
  --sam3-endpoint http://SAM3_HOST:SAM3_PORT \
  --planner claude_code --model claude-opus-4-7

Replace each *_HOST with the address of the machine running that service, and make sure it is reachable from the machine running RPent. Replace each *_PORT with the available port selected when starting the service. The three endpoint options are independent; when one is omitted, RPent starts that service on the current machine and selects an available port automatically. For all three services, omit the protocol to use HTTP, or use socket://HOST:PORT for socket RPC.