DAM is a detachable safety middleware that sits between any machine learning policy (or controller) and robot hardware. It intercepts every proposed action, evaluates it through a layered guard stack (L0–L3), and either passes, clamps, or rejects it — without modifying the policy weights or hardware drivers.

This design enables strong safety boundaries while keeping the learning/policy layer fully detachable and upgradable.

• 4-Layer Guard Stack: Progressive defense from perception (L0) → hardware (L3)
• Rust Data Plane: Deterministic, real-time-safe execution outside the Python GIL
• Stackfile-Driven: Swap hardware, policies, or safety rules via YAML. Zero Python code for simple tasks.
• Hot-Reload Boundaries: Update safety constraints without stopping the robot
• Fail-to-Reject: Any guard timeout, crash, or exception → immediate rejection
• MCAP Loopback Buffer: Capture ±30s of context around safety events for analysis
• Built-in Adapters: LeRobot (SO-ARM101) and ROS 2 support

Important Disclaimer: DAM is currently research and experimental-grade software. It is not certified for safety-critical or production use in human-collaborative or high-risk environments. Use at your own risk. We are actively working toward formal verification, worst-case timing analysis, and compliance-oriented documentation.

git clone https://github.com/ez945y/DAM.git
cd DAM
make setup

make run

After starting, open http://localhost:3000 in your browser and select a configuration template:

• Quick Start — Simulation only (no hardware needed)
• SO-ARM101 — Pre-configured for SO-ARM101 robot
• Custom — Create your own Stackfile

DAM acts as a transparent safety layer:

Policy / Controller
        │
        ▼
Proposed Action  ──────▶  [ Guard Stack L0–L3 ]  ──────▶  Validated Action
        ▲                       │      │      │                  │
        │                       │      │      │                  ▼
Observations & State  ─────────┘      │      └──────────▶  Fallback (Hold / Retreat / E-Stop)
                                       │
                                       ▼
                                 Decision: Pass / Clamp / Reject

Guard Layers

The final decision is the most restrictive outcome from all active layers.

v0.4.0 (Current focus)

• Runner-owned runtime lifecycle and control loop
• Image hub backed live preview and MCAP camera attachments
• ROS2 runner recheck and node preservation

v0.5.0

• More built-in boundary types
• Domain-specific bundles (manipulation, mobile manipulation, etc.)
• Extensive adversarial testing suite

Longer term

• Formal verification of critical safety paths
• Support for additional robot platforms
• Certification preparation artifacts

See CONTRIBUTING.md for details on:

• Setting up the development environment
• Code style and testing requirements
• How to propose new features or guard layers

We especially welcome help in the following areas:

• Safety testing and adversarial scenario development
• Real-time performance optimization
• Additional hardware adapters
• Documentation and example Stackfiles

DAM aims to make advanced robot safety modular, verifiable, and accessible to the embodied AI community.

Feedback and discussions are highly encouraged in GitHub Discussions.