policy_client: PolicyClient bridge, Face adapters, inference loop

[shantanuparab-tr]

Summary

Adds PolicyClient — the hardware/observer bridge that packs neutral Observations, drives one PolicyTransport over the async push/poll contract, and presents per-arm action slices through Face adapters to the existing teleop machinery. This is the chunk-timing core of the subsystem.

Stack: PolicyClient stack, based on #248. Review bottom-up (#252 → here).

Changes

• In include/trossen_sdk/hw/policy/policy_client.hpp + src/hw/policy/policy_client.cpp: PolicyClient (multi-inherits HardwareComponent + ObserverBase) and its per-arm Face. Implements the inference loop and its timing model:
  • timestep clock (⌊(now − epoch)·rate⌋) stamped on every observation and paired with the chunk's base_timestep on the return path;
  • drain-threshold θ firing gate — at θ=0 it packs at chunk exhaustion (openpi cadence, end-of-chunk pose); at θ>0 it fires early to overlap inference with playback;
  • freshness barrier so every observation snapshot is co-temporal (bounds cross-camera skew);
  • must_go starvation flag for queueing servers;
  • hold-last-action under any failure (Face::read() never throws), plus per-arm EMA output smoothing.
• In tests/test_policy_client.cpp + tests/CMakeLists.txt: drives the client against a FakeTransport (the two-arg constructor) — chunk apply/publish, hold-last-action, pause/resume, width-mismatch rejection.
• In cmake/policy_client.cmake: add policy_client.cpp to the gated sources.

Test Plan

• Builds cleanly with -DTROSSEN_SDK_ENABLE_POLICY_CLIENT=ON
• Tests pass — test_policy_client
• Lint passes (pre-commit run --all-files)
• Hardware — chunk-timing validated end-to-end via the live pi05 integration (see policy_client: stationary examples (openpi + lerobot) #255)

Breaking Changes

None. New gated sources behind TROSSEN_SDK_ENABLE_POLICY_CLIENT (default OFF).

Related

Based on #248. Stacked on #252; followed by #254–#255.

[shantanuparab-tr]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• policy_client: stationary examples (openpi + lerobot) #255
• policy_client: producer + hardware-registry integration #254
• policy_client: PolicyClient bridge, Face adapters, inference loop #253 👈 (View in Graphite)
• policy_client: transport interface, registry, and both transports #252
• policy_client: wire-format codecs (openpi msgpack + lerobot pickle) #251
• policy_client: neutral observation/action types #250
• policy_client: config schema + SdkConfig wiring #249
• control: per-episode hardware lifecycle driven by SessionManager #248
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.