LinkedIn Post Draft: DAM 0.5.0

I have been building DAM (Detachable Action Monitor) to make robot policy execution easier to inspect and safer to deploy.

In v0.5.0, a recorded dataset can be replayed through the same guard pipeline and then sent to a real robot only after validation. The live console and MCAP recording capture multiple cameras plus readable joint and motor telemetry, so a safety decision is reviewable rather than a black box.

The core design is lifecycle-aware: L0 checks unfamiliar observations before action, L1 constrains physical motion, L2 checks task-phase intent over consecutive cycles, and L3 watches hardware and host health during execution and fallback. Each layer shares the same runtime, event log, and replay path.

Highlights:

• guarded dataset-to-hardware replay for SO-101
• merged joint/telemetry inspection in live and recorded sessions
• multi-camera live preview and MCAP playback
• configurable consecutive-cycle reactions for task and hardware anomalies
• real experiment runners for evaluating safety behavior

This release made an important principle concrete for me: a robotics safety feature is only useful when the operator can understand why it acted.