Newton reading guide: what do the 35 papers study?

These works study physical computation, robot learning and simulation validation. Newton can be the engine being improved, or a tool for training, data generation or comparison.

This guide covers all 35 works in the complete bibliography, organized into seven topics. DexLab provides the topic grouping and plain-language summaries based on abstracts and author pages checked on 2026-11-02, not full-paper reviews or reproductions. The upstream distinction between 34 research entries and one comparison-only entry remains unchanged; external results do not count as DexLab experiment outcomes.

Start with a question

Question	Suggested route	What to look for
How is contact solved, and where do gradients come from?	Ostrich → FlashNeRD	The former studies hard-contact solving and implicit gradients; the latter keeps analytical collision detection and learns robot dynamics. Examine contact detection, force computation and differentiation separately.
Where does simulation diverge from reality, and how can it be adjusted?	GAUGE → Same Action, Different Outcome → RoboSimStudio	Read about measurement-based evaluation, repeated-trial variability, then model and parameter calibration.
How can world models support action selection?	PointCast → PileBelief	Separate future-motion prediction, interaction memory and downstream action selection. Prediction accuracy alone does not establish control success.

These routes suggest reading order, not paper-quality rankings.

Physics engines and contact solvers · 4

Paper	Research question
Ostrich	Use larger simulation timesteps under hard contact and friction, and compute gradients for action optimization.
Kamino	Simulate complex mechanisms with coupled joints, including kinematic loops, in parallel on a GPU.
Mixed Material Point Methods	Improve simulation of sand, snow and elastoplastic materials, including two-way interaction with rigid bodies.
Exact Reduced Coulomb Friction	Solve sticking and sliding while retaining the exact Coulomb friction law.

Evaluating and calibrating simulation · 7

Paper	Research question
GAUGE	Use real measurements to evaluate collision, friction and deformation in physics engines and video world models.
RoboSimStudio	Let people and coding agents adjust solvers, contact models and parameters to match observed physical behavior.
RigPI	Initialize physical parameters with a vision-language model, then refine them using motion, force measurements and differentiable simulation.
Informationally Decoupled Trajectory Design	Choose robot motions that separate the effects of different physical parameters for more reliable identification.
Same Action, Different Outcome	Measure variation across repeated cloth motions and test whether simulators reproduce that variation.
ContactBench 1.0	Provide reproducible physical contact tasks with matching simulation assets for evaluating robot manipulation.
Probabilistic Physics Digital Twin	Represent uncertain mass, friction and other properties as distributions to predict possible action outcomes and their likelihoods.

Learned dynamics and world models · 4

Paper	Research question
FlashNeRD	Keep analytical collision detection, learn faster robot dynamics and couple the robot bidirectionally to analytically simulated objects.
PointCast	Predict object point trajectories from robot actions. One architecture covers rigid objects, articulated objects, cloth and rope, with separate checkpoints per regime.
PileBelief	Retain physical evidence from previous excavation interactions to predict terrain changes and bucket volume.
PhysLDM	Generate volumetric deformation trajectories with latent diffusion. Newton is used only as a comparison baseline.

Cloth, rope and deformable manipulation · 6

Paper	Research question
ForwardDLO	Predict rope shape changes under bimanual actions quickly enough to evaluate more planning candidates.
RopeFormer	Retain action-response history across trials so a fixed policy can adapt to the current rope dynamics.
DefFlow	Jointly predict robot and object motion to learn grasping, moving, releasing and regrasping toward a target shape.
Learning Sim-Grounded Policies for Bimanual Rope Manipulation	Compare image observations with simulated rope state when learning from limited bimanual teleoperation demonstrations.
SIM1	Reconstruct and calibrate deformable simulation from limited real demonstrations, then expand synthetic training data.
High-Fidelity 4D Cloth Capture	Capture detailed cloth motion using multiple cameras and printed patterns, then complete occluded regions with physical constraints.

Robot control and reinforcement learning · 6

Paper	Research question
Backpropagating Through Simulation	Update control policies through simulation gradients and compare with sampled policy learning using PPO.
AlphaNIST	Learn grasping, reorientation and assembly in simulation using binary task rewards.
A Balanced Data Diet	Sample tasks near the policy capability frontier to reduce wasted experience in large-scale RL.
GM-Loco	Learn terrain-adaptive humanoid locomotion on yielding granular ground such as sand.
RoboMMT	Track expressive and interactive motions such as facing and pointing while maintaining dynamic balance.
Spectral Skills	Encode motion segments as skills that a planner can track, chain and compose for humanoid control.

Agents, environment creation and planning · 3

Paper	Research question
Gym-Anything	Use coding agents to create robotics simulations and audit their quality in an iterative loop.
GaP	Have agents assemble perception, planning and control modules into executable graphs and refine them through simulated rehearsal.
DetAug	Augment manipulation trajectories so execution can select routes around previously unseen obstacles.

Application studies · 5

Paper	Research question
LUMO	Design illuminated compliant finger pads whose optical response reveals contact location.
PersistGS	Use differentiable rigid-body simulation to track fully occluded objects in dynamic 3D reconstruction.
OrchardBench	Simulate apple harvesting, including branch bending, fruit detachment, foliage occlusion and plant damage.
RoboFin3D	Simulate robotic grinding and sanding with changes to workpiece geometry and surface appearance.
SceneFactory-3D	Hold traffic scenes and controllers fixed while changing road friction and grade to evaluate driving safety.

See the complete bibliography for full titles, authors, dates and upstream attribution.