Heavy NVIDIA simulation backends for strands-robots: Isaac Sim (USD + IsaacLab 3.0) and Newton/Warp (GPU-native, 4096+ envs, differentiable).

For the default lightweight MuJoCo backend, use strands-robots directly — both backends in this repo plug into the same Simulation AgentTool / SimEngine ABC and load via entry points, so the user-facing API is identical across all three.

Status: v0.2.0 is a re-scoped foundation release. The legacy SimEnv / SteppedSimEnv / GR00T policy code path moved upstream — see examples/MIGRATION.md. Backend code lands in v0.3.0+ (Isaac) and v0.4.0+ (Newton). Track the rollout in #8.

Authoritative sources: strands-labs/robots#96 (Newton design), strands-labs/robots#97 (Isaac Sim design).

• Pick MuJoCo for fast iteration, debugging, and macOS contributors.
• Pick Newton for fleet-scale RL training and differentiable workloads.
• Pick Isaac Sim for synthetic-data generation and photoreal teleop scenes.

System requirements: NVIDIA RTX GPU, Ubuntu 22.04+, CUDA 12+. macOS / Apple Silicon contributors should install strands-robots directly and skip this repo.

# Isaac Sim (~30 GB SDK download on first run)
pip install 'strands-robots-sim[isaac]'

# Newton + Warp
pip install 'strands-robots-sim[newton]'

# Both
pip install 'strands-robots-sim[all]'

strands-robots (and the default MuJoCo backend) are pulled in transitively.

Snippets are written against the post-R6 entry-point API. They become copy-paste-runnable when R6 / #13 and R7 / #14 (Isaac) or R11 / #18 (Newton) ship.

import strands_robots_sim                       # registers "isaac" / "newton" via entry points
from strands_robots.simulation import create_simulation

sim = create_simulation("isaac", rtx_mode="path_traced", headless=True)
sim.create_world()
sim.add_robot("so100")
sim.step(100)
frame = sim.render(camera_name="default")

import strands_robots_sim
from strands_robots.simulation import create_simulation

sim = create_simulation("newton", num_envs=4096, solver="mujoco")
sim.create_world()
sim.add_robot("so100")
sim.step(1000)
state = sim.get_state()                         # batched [4096, ...] tensors

The flagship — same LIBERO task on every backend, side-by-side success-rate + wall-time — lives in examples/libero_backend_matrix.py (filed as R15 / #22).

Per-backend examples:

• examples/libero_mujoco.py — default MuJoCo baseline (R5 / #12)
• examples/libero_isaac.py — RTX path-traced (R8 / #15)
• examples/libero_newton.py — single-env GPU (R12 / #19)
• examples/libero_newton_fleet.py — 4096 envs (R12 / #19)

Comparison skeleton (numbers committed as each variant lands):

Tracking umbrella: #8.

• Stage 1 — Foundation cleanup
  • R2 / #9 — remove legacy SimEnv / SteppedSimEnv / Libero env layer
  • R3 / #10 — remove duplicated GR00T policy / inference tool / tests / scripts / docs
  • R4 / #11 — README rewrite (this file)
• Stage 2 — MuJoCo LIBERO baseline
  • R5 / #12 — examples/libero_mujoco.py + examples/README.md
• Stage 3 — Isaac Sim (ships with v0.3.0)
  • R6 / #13 — entry-point backend registration
  • R7 / #14 — IsaacSimulation(SimEngine) backend
  • R8 / #15 — examples/libero_isaac.py
  • R9 / #16 — examples/isaac_replicator_synthdata.py
  • R10 / #17 — nightly GPU CI
• Stage 4 — Newton (ships with v0.4.0)
  • R11 / #18 — NewtonSimulation(SimEngine) backend
  • R12 / #19 — Newton LIBERO examples
  • R13 / #20 — examples/newton_diffsim_toy.py
  • R14 / #21 — extend nightly GPU CI for Newton
• Stage 5 — Flagship (ships with v0.5.0)
  • R15 / #22 — examples/libero_backend_matrix.py

Migrating from the 0.1.x API: examples/MIGRATION.md.

PRs welcome. hatch run lint (black / isort / flake8) and hatch run test (an import smoke until backend code lands) before submitting. Backend-specific tests will live under tests/isaac/ (R7) and tests/newton/ (R11). GPU CI is tracked separately in #17 (Isaac) and #21 (Newton).

Apache-2.0 — see LICENSE.

• strands-labs/robots — default MuJoCo backend, Simulation AgentTool, LIBERO adapter
• Strands Agents SDK
• NVIDIA Isaac Sim (IsaacLab)
• Newton (built on NVIDIA Warp)