vla-expert

name	vla-expert
description	Vision-Language-Action co-training with Wall-OSS-0.5, WALL-WM, and fleet simulation loops

VLA Expert Skill

When to use

• User wants zero-shot robot control via Wall-OSS-0.5 (flow-matching VLA)
• User wants synthetic training data via WALL-WM world model on Wan video prior
• User orchestrates worldlabs-mcp + robotics-mcp + avatarops for multiview episodes
• User runs DMuon co-training locally on exported shards

Workflow

1. vla_status - confirm upstream clones and dataset root
2. vla_fleet(operation='scenario_brief') - plan simulation boundary
3. Generate episodes with fleet MCPs; tag events (approaching, contact, lift, recovery)
4. vla_dataset(operation='ingest_episode') - register each run (include failures)
5. vla_dataset(operation='export_shard') - manifest for wall-x dataloaders
6. vla_training(operation='co_train_prepare') - DMuon + gradient-bridged co-train
7. vla_wall(operation='infer_prepare') - deploy to Raspbot / edge agents

Upstream

• wall-x: https://github.com/X-Square-Robot/wall-x (Wall-OSS-0.5 + WALL-WM + DMuon)
• X-VLA (lightweight alt): Tsinghua 0.9B flow-matching VLA for PEFT on edge

Env vars

Variable	Purpose
VLA_WALL_X_ROOT	Clone of wall-x
VLA_WALL_WM_ROOT	WALL-WM subtree (defaults to wall-x root)
VLA_DMUON_ROOT	DMuon training scripts root
VLA_DATASET_ROOT	Local episode + export store
VLA_CHECKPOINT	Model weights path
VLA_WORLDLABS_MCP_URL	worldlabs-mcp HTTP base
VLA_ROBOTICS_MCP_URL	robotics-mcp HTTP base
VLA_AVATAR_MCP_URL	avatarops HTTP base

Honesty

This MCP orchestrates and prepares; it does not run GPU training unless upstream is wired. Always report missing clones/checkpoints truthfully.