Summary
Proposes interleaving expensive remote/cloud VLA inference with a tiny (~7.4M parameter) local action predictor that uses current views, proprioception, and action history to fill in action chunks between cloud calls, cutting both latency and onboard power draw.
Key Contributions
- Ultra-lightweight local action predictor (ULAP) requiring no VLA hidden states or server round-trips once trained
- Demonstrates 48.8–76.7% reduction in cloud VLA calls while retaining 95–97.5% of baseline success rate across three sim benchmark/policy pairs
- Reports concrete power/latency numbers on Jetson Orin Nano vs. server-class GR00T inference
Strengths
- Addresses a genuine deployment bottleneck (communication latency + onboard compute/power budgets) that’s underexplored relative to model-capability research
- Independent training of the local predictor (no dependence on VLA internals) makes it broadly portable across cloud VLA backbones
Weaknesses
- Evaluated only in simulation for the call-reduction/success-rate tradeoff; real-world network latency and packet loss are not directly characterized
- Risk of compounding error during “local-only” intervals isn’t deeply analyzed for longer local-control stretches
Open Questions
- What is the failure mode when cloud connectivity is lost entirely for extended periods?
- How does the approach interact with safety-critical or contact-rich tasks where local mispredictions could be costly?
Significance
As VLA models grow toward billions of parameters, cloud/edge hybrid inference is likely to become standard practice for cost- and power-constrained robot fleets, making this a relevant systems contribution.