Krabby HAL Server - IsaacSim

HAL server implementation for IsaacSim with integrated parkour policy inference.

Overview

This package provides an entry point that runs both the IsaacSim HAL server and parkour inference client in the same process using inproc ZMQ for zero-copy communication.

Architecture

┌─────────────────────────────────────────┐
│         IsaacSim Process                │
│                                         │
│  ┌──────────────┐    inproc (ZMQ)     │
│  │ HAL Server   │◄──────────────────┐  │
│  │ (main thread)│                   │  │
│  └──────────────┘                   │  │
│         │                           │  │
│         │ publishes observations    │  │
│         │ receives commands         │  │
│         │                           │  │
│  ┌──────────────────────────────────┴─┐│
│  │ Parkour Inference Client           ││
│  │ (separate thread)                  ││
│  │  - Polls observations              ││
│  │  - Runs policy inference           ││
│  │  - Sends joint commands            ││
│  └────────────────────────────────────┘│
└─────────────────────────────────────────┘

Installation

From source (development)

cd hal/server/isaac
pip install -e .

With optional dependencies

pip install -e ".[dev]"

Usage

Command line

After installation, use the krabby-hal-server-isaac command:

krabby-hal-server-isaac \
  --task Isaac-Anymal-D-v0 \
  --checkpoint /path/to/model.pt \
  --action_dim 12 \
  --obs_dim 753

Python module

python -m hal.server.isaac.main \
  --task Isaac-Anymal-D-v0 \
  --checkpoint /path/to/model.pt \
  --action_dim 12 \
  --obs_dim 753

Arguments

Required:

• --task: IsaacSim task name (e.g., Isaac-Anymal-D-v0)
• --checkpoint: Path to model checkpoint file
• --action_dim: Action dimension (typically 12)
• --obs_dim: Observation dimension (typically 753)

Optional:

• --inference_device: Device for inference, cuda or cpu (default: cuda)
• --observation_bind: Observation endpoint (default: inproc://hal_observation)
• --command_bind: Command endpoint (default: inproc://hal_commands)
• --backend: direct (default) or fullsim (three-board host firmware on the hex env)
• --firmware-transport: with fullsim, in_process (default) or serial_sdk (KrabbyMCUSDK over PTY)

Plus all AppLauncher arguments for IsaacSim configuration.

Fullsim firmware transport (crab hex)

--backend fullsim runs three-board host firmware on the hex scene (18 DOFs). The krabby-isaacsim entrypoint (run_hal_server.sh) adds --backend fullsim automatically for crab hex launches (--usd, --robot hex, or a CrabHex task) when you omit --backend. Pass --backend direct to force the old Isaac-only path on hex. Default --firmware-transport in_process is usually enough; use serial_sdk only when you need the Jetson/GUI pyserial path in the loop.

python -m hal.server.isaac.main --task <crab-hex-task> --backend fullsim \
  --firmware-transport in_process …

Jetson production opens real USB with KrabbyMCUSDK; it does not use these flags. serial_sdk on Isaac uses sparse B batches on the PTY shim; Jetson still sends full 18-joint batches on hardware USB. PTY setup for GUI/scripts: firmware/sim/VIRTUAL_SERIAL.md. Isaac firmware pytest: ./images/isaacsim/run_firmware_isaac_pytest.sh (mounts live hal/ / compute/).

Components

HAL Server (hal.server.isaac.IsaacSimHalServer)

• Extracts observations from IsaacSim environment
• Publishes observations via ZMQ PUB socket
• Receives joint commands via ZMQ PULL socket
• Applies commands to simulation

Parkour Inference Client (compute.parkour.inference_client.ParkourInferenceClient)

• Runs in separate thread
• Polls observations from HAL server
• Runs parkour policy inference
• Sends joint commands back to HAL server

Development

Project Structure

hal/server/isaac/
├── pyproject.toml      # Package configuration
├── README.md           # This file
├── __init__.py         # Package init
├── main.py             # Entry point with integrated inference
└── hal_server.py       # IsaacSimHalServer implementation

compute/parkour/
├── inference_client.py # Parkour inference client
├── policy_interface.py # Policy model interface
├── types.py            # Parkour-specific types
└── mappers/            # HW ↔ Model mappers

Running Tests

pytest tests/

Notes

• This package uses inproc ZMQ by default for same-process communication (zero-copy, high performance)
• For distributed deployment (server and client on separate machines), use TCP endpoints instead
• The parkour inference client runs on a separate thread to avoid blocking the simulation loop