NEWTONTERRAIN LAB

OPEN SOURCE · HUMANOID SIMULATION

Give your humanoid
new ground.

Move beyond flat-floor tests. Build mixed terrain, connect your simulator, and see how a motion-tracking policy responds.

Experimental research tool · Newton MPM + Isaac Lab / MuJoCo

ARCHIVED FROZEN-TEACHER ROLLOUTReconstructed MPM surfaces · recorded replay
167.5 m

across the default map

256

exclusive terrain tiles

15

material combinations

4

curriculum levels

LATEST STUDY / REPORT UPDATED 21 SEP 2026

Training completed.
Stable tracking is still the goal.

Two matched 256-rollout experiments tested shallow-to-deep practice against direct training. Both missed the original-depth stability and tracking requirements.

Study complete · No checkpoint promoted · Research baseline retained

Starting policy56soil task failures
2 → 6 → 14 cm curriculum535.36% fewer; acceptance failed
Direct 14 cm training543.57% fewer; acceptance failed

Original 14 cm soil · Three development motions · Seed 23 · 30 seconds per terrain family, with automatic resets. Neither candidate retained mud tracking accuracy or completed full references in every soil family.

The experiment

Frozen SONIC + a learned leg residual. Four environments, 50 Hz control, 2 cm MPM grid, original motion timing.

The failure analysis

Large pelvis tilt, tracking drift and difficult foot extraction. Reward conflict is one hypothesis; physics and practice distribution also matter.

The evidence

All twelve reward terms, PPO and actuator settings, soil parameters, reset events, source hashes, interrupted attempts and measured results.

01 / IN THE SIMULATOR

Watch the policy meet the terrain.

Archived 13 September frozen-teacher recordings at a 4 cm grid. These are not videos of the latest trained policies. Failures and resets remain visible.

Starts on sand + mud + rocks. Curriculum demotion and episode resets remain visible.

MODEL & MEASUREMENT

SONIC teacher / step 8,000

The strongest measured teacher in the available local checkpoint comparison. This is a frozen policy evaluation, not a terrain-trained success claim.

Video
10 s · 25 fps
Recorded env terminations
See record
Physics grid
4 cm · exploratory
New native runs
Receipts loading…
View the measurement record ↗

02 / ONE MAP, MANY CONDITIONS

Explore your next experiment.

A connected rigid map with localized material patches. Select a tile to inspect its composition and curriculum level.

GENERATED MAP / SEED 23

Loading the generated map…

GroundSandMudRocks

03 / DESIGNED TO CONNECT

Keep your simulator.
Add material interaction.

One shared terrain description connects rigid geometry, material simulation and motion-tracking evaluation. Start with exports; attach the live bridge when you need deformable contact.

01

Generate

Seeded terrain JSON
Rigid USD + MuJoCo XML

02

Couple

Newton advances materials
Your engine advances the robot

03

Evaluate / train

Per-env resets & curriculum
Policy metrics + surface replay

04 / FROM DEMO TO YOUR TASK

Start small. Build on evidence.

Map generation works on CPU. Live materials require the Newton GPU worker; native policy runs additionally need a compatible Isaac/SONIC installation and your model assets.

CPU-only geometry export. No teacher checkpoint or Isaac installation required.

Read the full integration guide ↗
GENERATE / EXPORT

Choose the integration depth you need
PlatformReady to useWhat you bring
NewtonMPM materials, probes, surface reconstructionCompatible CUDA GPU
MuJoCoRigid XML + live external-wrench bridgeMatching robot model and policy adapter
Isaac Sim / Isaac LabShared USD, exclusive origins, resets, surfacesNative installation and compatible task
SONICFrozen-policy eval + PPO warm startCompatible TRL fork, weights and motion data

RESEARCH SOFTWARE, WITH RECEIPTS

Useful today.
Clear about the limits.

Read the physics assumptions ↗

Demonstrated

  • Native two- and four-environment execution
  • Selective resets preserve peer states
  • Continuous surfaces from measured MPM samples
  • Two completed 256-rollout adaptation arms with numerical checks

Still to establish

  • Sand and mud calibration against real measurements
  • Reliable deep-soil tracking and useful learning gains
  • High-throughput training at large batch sizes
  • General adapters beyond the tested SONIC fork

Water is excluded. Smooth surfaces improve the display, not the physics model. Importing USD or XML alone does not activate deformable materials.

MAKE THE NEXT TEST MORE INFORMATIVE

Your policy. A less forgiving floor.

Explore the open-source toolkit