robot-dam 0.6.0

Real-time safety guard pipeline for robot arms — clamp, reject, or fallback on unsafe actions.

DAM — Detachable Action Monitor

See where your policy breaks — before your hardware does.

Install · Quick Start · Isaac Sim · Architecture · Docs

https://github.com/user-attachments/assets/a10711ea-a419-4aee-ba06-de1e2d437d49

Install

pip install robot-dam            # includes torch, lerobot, mcap, dam-rs (Rust extension)
pip install robot-dam[isaac]     # + Isaac Sim 6.0 adapter
pip install robot-dam[full]      # + ROS2, REST API, QP solver, everything

Or from source:

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

---

What It Does

ML policies always output the next action — they have no idea when something is unsafe. DAM intercepts every action and evaluates it through a layered guard pipeline before it reaches hardware. Each action is either passed, clamped (adjusted to stay safe), or rejected.

The goal isn't to hide failures. It's to make them visible and easier to understand.

Important: DAM is experimental research software, not a certified safety system.

---

Quick Start

git clone https://github.com/ez945y/DAM.git
cd DAM
make setup    # Python venv + Rust extension + npm install (~3 min)
make run      # Backend :8080 + Console :3000

Open http://localhost:3000 — the demo Stackfile replays a dataset through the full guard pipeline without hardware.

CLI

.venv/bin/dam doctor                                # check environment
.venv/bin/dam callbacks                             # list 18 built-in safety checks
.venv/bin/dam validate examples/stackfiles/*.yaml   # schema-check Stackfiles
.venv/bin/dam run examples/stackfiles/demo.yaml --cycles 200 --task demo

Make Targets

Command	What it does
make setup	First-time install (venv + Rust + npm)
make run	Backend + pre-built frontend
make dev	Backend + frontend with hot-reload
make test	Full test suite (688+ Python tests + 109 frontend tests)
make record	Safe IL recording with DAM guards
make callbacks	List all 18 built-in safety checks
make validate	Validate example Stackfiles

---

Safe Recording for Imitation Learning

During data collection, DAM smooths motions and catches bad data before it enters your dataset. Integrates with LeRobot — one file configures hardware, safety boundaries, and recording.

make record                                # reads examples/stackfiles/safety.yaml
make record ARGS="--dataset.num_episodes=20"   # CLI args override YAML

Python API (3 levels)

import dam

# Level 1: one-liner (notebooks)
safe_action = dam.safe(action, obs, stackfile="safety.yaml")

# Level 2: stateful guard (recording loops + Isaac Sim)
guard = dam.SafetyGuard("safety.yaml", task="record")
safe_action = guard(action, obs)  # dict → dict, ndarray → ndarray, Tensor → Tensor

# Level 3: lerobot pipeline (one line addition)
from dam import SafetyProcessorStep
robot_action_processor.steps.insert(0, SafetyProcessorStep("safety.yaml"))

The recorded dataset contains only safe actions — your policy trains on data that already respects physical constraints. See Safe Recording Guide for full details.

---

Isaac Sim Integration

DAM works as a drop-in safety filter for Isaac Sim 6.0+ control loops. Accepts and returns torch.Tensor on any device — no CPU↔GPU copies needed.

import dam
from dam.adapter.isaac import IsaacSafetyFilter

# Single env — SafetyGuard accepts torch.Tensor directly
guard = dam.SafetyGuard("franka_safety.yaml", task="manipulation")
safe_action = guard(action_tensor, obs_tensor)  # same device/dtype

# Multi-env (Isaac Lab) — vectorized filter
filt = IsaacSafetyFilter("franka_safety.yaml", task="manipulation", num_envs=4096)
safe_actions = filt(actions_batch, obs_batch)    # (N, J) → (N, J)

# Auto-detect from ArticulationView
filt = IsaacSafetyFilter.from_articulation(franka_view, "franka_safety.yaml")

Bundled presets: so101_follower, franka_emika_panda. See examples/isaac_franka_demo.py for the full demo.

---

Architecture

Guard Layers

Layer	What it checks	Example checks
L0 OOD Detection	Is the observation in-distribution?	Real-NVP / Memory Bank anomaly scoring
L1 Physical Kinematics	Can this move physically happen?	Joint limits, workspace bounds, velocity caps
L2 Task Execution	Does this make sense for the task?	Gripper sequence, progress enforcement
L3 Hardware Monitoring	Is the hardware healthy?	Temperature, current, voltage, heartbeat

L0 and L1 run in parallel. The final decision is the most restrictive outcome across all active layers.

Stackfile Configuration

Everything is configured in a YAML Stackfile:

boundaries:
  joint_position_limits:
    layer: L1
    type: single
    nodes:
      - callback: joint_position_limits
        params:
          upper: [1.8243, 1.7691, 1.8326, 1.8067, 3.0741, 1.7453]
          lower: [-1.8243, -1.7691, -1.8326, -1.8067, -3.0741, 0.0]

Start from examples/stackfiles/minimal.yaml or see the Stackfile Walkthrough.

18 Built-in Safety Checks

Run dam callbacks to list them all:

• L0: OOD detector (Real-NVP, Memory Bank, Welford)
• L1: Joint position/velocity limits, workspace bounds, keep-out zones, orientation, geofence, Cartesian velocity, smoothness
• L2: Task gripper sequence enforcement
• L3: Temperature, current, voltage, force/torque, hardware watchdog, host health

---

Documentation

Goal	Start here
Learn DAM step by step	Learn DAM
Read a Stackfile	Stackfile Walkthrough
Common config edits	Common Stackfile Edits
Understand the console	Console Walkthrough
Prepare for hardware	Hardware Readiness
Troubleshooting	Troubleshooting

---

Contributing

See Contributing. We welcome safety testing, hardware adapters, performance optimization, and example Stackfiles.