TroubleshootingΒΆ

  • Dataset URL is a placeholder: supply the published URL/checksum, or unpack a local bundle. dw run requires the selected scene files to be installed.

  • Out of GPU memory: start with one robot and Frontier/Lawnmower, check other GPU processes, or assign model services to another available GPU. For CoNavGPT, an existing VLM API can replace the local VLM.

  • Existing checkout has changes: setup preserves tracked edits. Use the managed .runtime installation or a separate clean supported checkout; do not point setup at the research fork.

  • Endpoint connection fails: the URL is resolved from containers. Check the VLM API base URL, configured model name, API key, and reachable host/IP.

  • No odometry, takeoff rejected, or clock stalled: inspect container.log and policy logs in the run directory. Failed missions are not scored as completed episodes.

  • Missing USD textures/references: the public bundle must be self-contained. Native USD dependency validation is required before release.

  • PPL unavailable: no validated reference route matches this discovery subset, fleet size and starts. TIP and Progress can still be computed for a completed episode.

  • Already running project: use ./dw down, or a different project name and nonconflicting subnet. Existing research containers are separate from this runtime.

Reach out or raise an issue if you encounter any problems.