ShardVLA

ShardVLA is a lightweight glue layer around LeRobot policies for running a sharded visual-language-action (VLA) stack:

Running the high and low nodes

1. High node (gRPC server)

Run this in terminal 1:

cd /home/getpanked/projects/ShardVLA

python -m shardvla.cli.run_high_node \
  --port 50051 \
  --log-level INFO \
  --model-path aractingi/bimanual-handover-groot-10k  # or a local path

Flags:

• --port: gRPC port to listen on (default 50051).
• --log-level: logging level (DEBUG, INFO, WARNING, ...).
• --model-path: HuggingFace repo ID or local path for the GR00T model.

Note: in offline / restricted environments you will need to point --model-path at a local GR00T checkpoint instead of a remote HF repo.

2. Low node (runtime bridge)

Run this in terminal 2:

cd /home/getpanked/projects/ShardVLA

python -m shardvla.cli.run_low_node \
  --server-addr 127.0.0.1:50051 \
  --robot-id robot1 \
  --session-id sess-001 \
  --log-level INFO

---

Using ShardVLA as a library

You can embed ShardVLA components directly in your own code:

from shardvla.runtime import ShardVLARuntime
from shardvla.transport.grpc_intent_server import IntentServer
from shardvla.policies.groot import GrootPolicyAdapter, GrootPolicyConfig
from lerobot.policies.groot.processor_groot import make_groot_pre_post_processors


# High node wiring
backbone = GrootPolicyAdapter(GrootPolicyConfig(model_path="your/model/path", strict=False))
preprocessor, _ = make_groot_pre_post_processors(config=backbone.config, dataset_stats=None)
server = IntentServer(backbone=backbone, preprocessor=preprocessor, port=50051)

# Low node wiring
runtime = ShardVLARuntime()

This separation lets you swap PolicyAdapter implementations (e.g., dummy policies for testing) without touching the transport or runtime code.