Balance board

Articulations

Restore balance by placing the missing counterweight.

Play0.00 s

¼×½×1×2×

Loop

The steel preset (the default), recordedon an NVIDIA RTX PRO 6000 Blackwell, sped up 2× (the readout shows sim time).

Run it

cd examples
# the default preset (steel), then the others
uv run shelf play balance_board
# change solver settings live
uv run shelf tune balance_board

Source: examples/shelf/07_articulations/balance_board/ · ported from steel-webgpu assets/scenes/balance_board (hero_montage export) @ ad03e49

Presets

Preset	Solver	Status	Checked
steeldefault	steel	works	2026-11-10, newton f6fc82a8b, NVIDIA RTX PRO 6000 Blackwell Server Edition

Solvers with no row are untested on this example. The full settings are inpresets.toml.

Code

example.py

examples/shelf/07_articulations/balance_board/example.py

# SPDX-FileCopyrightText: Copyright (c) 2026 Skild AI and the Steel contributors
# SPDX-License-Identifier: Apache-2.0
# Direct source: steel-webgpu; see presets.toml and THIRD_PARTY.md.
# Modified for Steel Shelf: preset-driven execution and shared runner integration.

###########################################################################
# Example Balance board
#
# A spring-loaded balance beam tips because one counterweight is missing; the arm puts it back and the beam levels out.
#
# Rebuilt from steel-webgpu's exported scene (assets/scenes/balance_board: scene.json
# and scene.bin, copied next to this file); see common/exported_scene.py. The
# visual meshes draw when a steel-webgpu checkout sits next to the shelf.
#
# Command: python -m shelf run balance_board
#
###########################################################################

import numpy as np

import newton.examples
from common.exported_scene import ExportedSceneExample


class Example(ExportedSceneExample):
    file = __file__

    def test_post_step(self):
        pivot = float(self.state_0.joint_q.numpy()[self.model.joint_q_start.numpy()[16]])  # balance.pivot.q
        self._tip = min(getattr(self, "_tip", 0.0), pivot)
        self._pivots = (getattr(self, "_pivots", []) + [pivot])[-60:]

    def check(self):
        """The beam tips (about -0.28 rad), the arm places the missing weight and the beam levels out at +0.023 rad
        (a converged 16 x 32 run) and stays there."""
        assert self._tip < -0.2, f"the beam only tipped to {self._tip:.3f} rad"
        level = self._pivots[-1]
        assert abs(level - 0.023) < 0.01, f"the beam ends at {level:.4f} rad, converged +0.023"
        assert np.ptp(self._pivots) < 0.002, f"the beam still moves {np.ptp(self._pivots):.4f} rad in the last second"


if __name__ == "__main__":
    parser = Example.create_parser()
    viewer, args = newton.examples.init(parser)
    newton.examples.run(Example(viewer, args), args)

example.py on GitHub

presets.toml

examples/shelf/07_articulations/balance_board/presets.toml

example = "balance_board"
source = "steel-webgpu assets/scenes/balance_board (hero_montage export) @ ad03e49"
default = "steel"

[check]
# Arguments `python -m shelf check` passes in test mode: 720 frames, the whole task and two still seconds (the beam levels out by frame ~600); the example's check() asserts the outcome.
args = ["--num-frames", "720"]

[gallery]
short = "Restore balance by placing the missing counterweight."
title = "Balance board"
frame = 420

[xpbd]
status = "untested"

[vbd]
status = "untested"

[mujoco]
status = "untested"

[featherstone]
status = "untested"

[semi_implicit]
status = "untested"

[feather_pgs]
status = "untested"

[steel]
status = "works"
# 2 substeps x 8 iterations, immediate response, contact warm start (so no sleeping: this Newton's Steel does
# not combine them): the fastest setting that matches a converged 16 x 32 run with the palm-on-weight grip under
# 2 mm. The beam tips to -0.276 rad, the arm places the missing weight within 1-7 mm of where the converged run
# puts it, and the beam settles at +0.024 rad (converged +0.023) by frame 568. With the gripper's finger mimic
# (cfeff10) 2 x 4 grips 2.8-3.0 mm deep (1.8 mm without the mimic); 2 x 8: 1.6-1.9 mm, 3 x 8: 0.9 mm at 3.6 ms.
# Without warm start 2 x 4 reaches 5.2 mm. Every setting tried does the task. Velocity limits off: every exported
# limit is 1e6. Rows: at most 54 used.
sim = { fps = 60, substeps = 2 }
solver = { solve_mode = "matrix_free", iterations = 8, articulated_contact_response = "immediate", enable_joint_limits = true, joint_limit_activation_gap = 0.1, enable_joint_velocity_limits = false, contact_warmstart = true, articulation_constraint_max = 128, free_body_constraint_max = 128 }
checked = "2026-11-10, newton f6fc82a8b, NVIDIA RTX PRO 6000 Blackwell Server Edition"

[kamino]
status = "untested"

[implicit_mpm]
status = "untested"

[style3d]
status = "untested"

presets.toml on GitHub

← Loop dr legsAll examplesBox turn →