Dexsuite Kuka Allegro Lift Task (Newton)#

This example is an experimental showcase for the Isaac Lab 3.0 Newton physics backend, demonstrating dexterous object lifting with the Kuka Allegro hand. Training is performed in Isaac Lab and the resulting checkpoint is evaluated in Arena — both using Newton (MuJoCo-Warp solver) for physically accurate contact modelling during dexterous manipulation.

Important

Newton Physics — Experimental

All Arena environments can switch to Newton physics by passing --presets newton on the command line (mirrors Isaac Lab’s presets=newton Hydra override). However, Newton support is experimental — only the dexsuite_lift example has been verified to work with Newton under the current simulation settings. Other environments may require additional tuning of solver parameters and physics parameters to run correctly using Newton physics.

Task Overview#

Task ID: dexsuite_lift

Task Description: The Kuka arm with an Allegro dexterous hand lifts a procedurally generated cuboid to a commanded target position using joint-space actions and contact-rich proprioceptive observations — including fingertip contact forces, hand-tip body states, object point cloud, and 5-step observation history.

Key Specifications:

Property	Value
Tags	Dexterous manipulation, contact-rich
Skills	Reach, Grasp, Lift (multi-finger)
Embodiment	Kuka LBR iiwa + Allegro Hand (7 DOF arm + 16 DOF hand)
Scene	Procedural table (static background) with ground plane and lighting
Objects	Procedural lift cuboid (`procedural_cube`)
Policy	RSL-RL PPO (`DexsuiteKukaAllegroPPORunnerCfg`)
Training Method	Reinforcement Learning (on-policy PPO) — trained in Isaac Lab
Physics Backend	PhysX (default) or Newton (`--presets newton`)
Simulation Rate	200 Hz physics, 50 Hz control (decimation = 4)
Episode Length	6 seconds
Closed-loop	Yes (50 Hz control)
Command Space	Target position [x, y, z], position-only, resampled every 2–3 s

Note

The physics backend defaults to PhysX. Pass --presets newton to policy_runner.py (Arena) or presets=newton to train.py (Isaac Lab) to switch to Newton (MuJoCo-Warp solver), which provides more physically accurate contacts for dexterous manipulation at the cost of slower simulation.

Workflow#

This tutorial covers training in Isaac Lab and evaluating the resulting checkpoint in Arena.

Prerequisites#

Start the Isaac Lab Arena docker container:

./docker/run_docker.sh

Workflow Steps#

Follow the steps below to complete the workflow: