[geyan21]

Stacked on #6 — this branch is based on it, so the diff includes its commits until it merges. The rd infer changes proper are the last two commits.

Summary

Adds rd infer, a local inference loop for deploying a trained policy on the robot host, with an optional HG-DAgger interactive-correction mode (--intervene): the policy runs autonomously, the operator grabs a leader arm the moment it starts to fail, corrects, and hands back — and the whole rollout is recorded as one episode with a per-frame control_source flag (consumed by the convert/shardify plumbing from #6). Follows the ABC recipe: record whole rollouts and mix/upweight corrections at train time, rather than training on interventions alone (see docs/guide/recovery.md).

Why a new command instead of extending rd serve

rd serve is a remote policy server (chiral protocol): fire-and-forget streaming, leaders off, hard-estop on disconnect. HG-DAgger needs a per-step rendezvous between observation, action, and takeover state, plus powered leader arms — a different control topology. So the loop lives in a new raiden/inference.py hosting the model in-process behind a small ModelBridge ABC (load / reset / predict), while reusing RaidenPolicyServer's cameras, depth, proprio, and motor control. The hooks added to server.py are default-off, so rd serve behavior is unchanged.

Design highlights

• Instant, seam-free takeover — during autonomy the powered leaders shadow the followers' measured pose, so pressing a trigger hands you the arm exactly where it is; the correction is a joint-space delta (leader and follower are kinematically identical YAM arms), giving a zero seam by construction. Handback re-observes and re-predicts from the pose you left.
• Bumpless gripper clutch — the leader trigger is unmotorized and can't be shadowed; the policy's grip command is held until your trigger crosses it, then tracks absolutely.
• Safe shutdown ordering — Ctrl+C homes and de-energises the arms before the heavy save (SVO2 flush + np.savez_compressed); the reverse order starves the 100–250 Hz motor threads past the DM4310 watchdog and drops CAN mid-motion.
• Recording shares the live ZED handles — one camera feeds the policy and records SVO2 simultaneously; recording toggles quiesce the grab loop first (a concurrent grab() + enable_recording() deadlocks the ZED SDK).
• Single leader-bus owner — all leader reads/writes stay on one input thread; the policy-runaway estop path stops that thread before commanding the leaders.
• The per-step joint-delta limit (--max-joint-delta) applies in policy mode only — a deliberate human correction is not a fault.
• Action convention is left-then-right ([left(6), l_grip, right(6), r_grip]), matching rd serve and ee_pose IK output.

❓ Question for maintainers: leader trigger calibration

On our rig the leader trigger's physical throw saturates a few percent short of the passive encoder's hardcoded range, so a full squeeze only drives the gripper command down to ~0.04 instead of 0.0 — the follower never fully closes, holding a visible gap and dropping small parts. That is invisible in teleop but bites in HG-DAgger, where a takeover exists precisely to fix a failed grasp.

_grip_clutch therefore snaps a trigger within _GRIP_FULL_CLOSE_SNAP (0.05) of the bottom to a complete close. It is deliberately confined to the takeover path, not the shared YAMLeaderRobot.get_info() where it would also change rd teleop / rd record for every user based on our rig's calibration.

If your triggers reach a true full squeeze, this snap is a harmless no-op on the bottom 5% and you may prefer to drop it — happy to remove it or make it a flag. Worth confirming on your hardware either way, since the underlying saturation would affect teleop grasps too.

Usage

rd infer --bridge my_model.bridge:MyBridge --ckpt-path /path/to/ckpt \
         --intervene --dagger-task box_fold_dagger_r1 --dagger-instruction "fold the box flap"

Full operator guide (takeover loop, round organization for the 80:10:10 mix, writing a ModelBridge): docs/guide/recovery.md.

Status / test plan

tests/test_controller.py (new) covers the takeover control logic hardware-free — 9 tests over _grip_clutch and joint_delta_takeover: the clutch holds the policy's grip until the passive trigger crosses it, never moves the gripper in an unintended direction at takeover, latches through a reversal, reaches a full close via the snap, and joint_delta_takeover is checked for the zero-delta takeover edge and for an un-grabbed arm holding its grip. This is the repo's first tests/ directory; no test job exists in CI yet, so they run locally via uv run pytest tests/ -v.

⚠️ Draft. The offline logic is reviewed and the data contract with convert/shardify is covered by the tests referenced in #6, but the hardware loop needs on-robot validation before this is merge-ready:

• takeover / handback seam (no jump) in joint and ee_pose modes
• leader shadow tracking + gripper clutch feel
• Ctrl+C ordering: home → de-energise → save, arms never go limp
• SVO2 start/stop against live grab loops (no deadlock, no dropped frames)
• runaway-delta estop path with the leader input thread active
• recorded episode round-trips through rd convert / rd shardify --keep interventions

ruff format --check + ruff check clean; mkdocs build --strict clean with the docs-CI dependency set.

🤖 Generated with Claude Code