Imagine you're writing a recipe book, but every existing cookbook assumes you already know French technique, or Italian technique, or molecular gastronomy — and each uses different names for the same ingredients. No single book starts from flour and walks you through all three traditions in the same measuring system. That's the state of contact-dynamics literature for roboticists, and this tutorial is the first attempt to be that unified cookbook. The committed claim: no self-contained pedagogical treatment exists that brings complementarity-based contact formulations and their solvers together for robotics audiences, with complete derivations and unified notation. This isn't a new algorithm or a new result — it's a bridge document connecting rigid-body dynamics, conic optimization, and frictional contact theory that have lived in separate literatures (computational mechanics, computer graphics, individual simulator docs) for decades. The authors — spanning MPI Tübingen, UPenn, INRIA, and Laas-CNRS — are placing a bet that the field's bottleneck is pedagogical, not algorithmic. The architectural backbone is the cone complementarity problem (CCP). Contact forces live in friction cones; complementarity conditions enforce that surfaces don't interpenetrate and that friction dissipates energy correctly. The tutorial walks through how these conditions encode three contact modes — separation, sticking, sliding — without the simulator needing to prescribe which mode is active. This is the key insight: the physics self-selects the contact regime through the optimization structure, rather than requiring a mode schedule. The paper distinguishes exact transcriptions from convex relaxations, and is unusually honest about the latter: relaxations can introduce artifacts like spurious lift-off (objects floating above surfaces) and creep (objects sliding when they should stick). On the solver side, the tutorial covers both first-order methods (projected gradient, ADMM-style splitting) and Newton-based methods, explaining regularization and proximal strategies. This is where the practical value lives for roboticists building their own simulators or trying to understand why MuJoCo, Drake, or Pinocchio behave differently on the same scene. The paper includes 14 algorithms — pseudocode you can actually implement — and 3 comparison tables that map formulation choices to simulator behavior. The existence-and-uniqueness discussion is particularly valuable: when multiple valid contact force solutions exist for one configuration, the solver picks one, and different solvers pick differently. Understanding this is essential for anyone debugging sim-to-real transfer. The integrity picture is unusual for a tutorial: there are no benchmarks to game, no SOTA to beat. The validation is structural — do the derivations connect correctly, do the algorithms reproduce known results, do the comparisons to existing simulators hold? The 9 figures and 3 tables serve as worked examples rather than experimental evidence. The author list includes maintainers of Pinocchio (Carpentier) and researchers deeply embedded in the contact-dynamics community (Posa, Toussaint, Martius), which provides credibility but also means the tutorial reflects their particular vantage point. The practical milestone isn't a number — it's adoption. The field needs this tutorial to become the default onboarding document so that every new PhD student or engineer building a contact simulator starts from the same mathematical foundation rather than reverse-engineering one simulator's source code. The real test is whether this displaces the current patchwork of Stewart-Trinkle (2000), Anitescu-Potra (1997), and individual simulator documentation as the entry point. If it does, the downstream effect on differentiable simulation and sim-to-real transfer could be substantial, because practitioners will finally understand what their tools are actually doing under the hood.
The analogy
Every existing cookbook for contact dynamics assumes you already know French, Italian, or molecular technique — and each uses different ingredient names. This tutorial is the first to start from flour and teach all three traditions in one measuring system.
Unified cookbook
One book teaching French, Italian, and molecular techniques from shared fundamentals in consistent measurements
Contact tutorial
One document teaching mechanics, graphics, and simulator-specific contact formulations from shared math in unified notation
Honest assessment
This is the most valuable kind of academic contribution that gets the least credit: a bridge document that makes an entire subfield accessible by synthesizing scattered knowledge into one coherent treatment. It won't win best-paper awards because it proposes nothing new, but it may do more for the field's velocity than any single algorithmic advance this year. The test is adoption — if this becomes the standard onboarding reference, it's a force multiplier. If it stays a one-off arXiv upload, it's a well-written document that six people cite.
What Is CLAIMS?
70
/ 100
BedrockNews applies a structured analytical framework called CLAIMS to academic papers and technical research releases.
How CLAIMS worksClaim
4/5
Ladder
3/5
Architecture
5/5
Integrity
3/5
Milestone
2/5
Successor
4/5
What is this paper actually committing to?
Genuine gap identified: no unified pedagogical treatment of complementarity contact for robotics exists. The claim is about synthesis, not novelty — honest and correctly scoped for a tutorial paper.
The field fight
Is the robotics contact-simulation bottleneck algorithmic (we need better solvers) or pedagogical (we need practitioners to understand existing formulations)?
Side A
Better algorithms and faster solvers are the priority; practitioners can learn on the job
Held by: MuJoCo, Drake, Isaac Sim development teams
Side B
The formulations exist but are scattered and inconsistent; unified pedagogy would accelerate the whole field
Held by: This paper, Carpentier (Pinocchio), Posa, Toussaint
Milestone tracker
Unlocks: Runnable, comparable implementations of all major contact formulations on shared benchmark scenes
Current
14
algorithms
this tutorial
Next
1
reference codebase
companion implementation
~1-2 years if authors or community build it
Verified Facts (7)
32 pages, 9 figures, 3 tables, 14 algorithms
Authors: Onur Beker, A. René Geist, Michael Posa, Marc Toussaint, Justin Carpentier, Georg Martius
Categories: cs.RO, cs.GR, math.OC
Covers cone complementarity formulations, distinguishes exact transcriptions from convex relaxations
Relaxations can introduce artifacts such as lift-off and creep
Covers first-order and Newton-based methods, regularization and proximal strategies
Connections to widely used simulators, optimal control, and differentiable simulation
Practical Implications (4)
Roboticists can now understand WHY MuJoCo and Drake produce different contact forces on the same scene — different solver choices among multiple valid solutions
The 14 pseudocode algorithms are directly implementable for anyone building custom contact simulators or debugging sim-to-real transfer
Understanding convex relaxation artifacts (lift-off, creep) lets practitioners diagnose simulation bugs that look like physics bugs
Differentiable simulation users gain clearer understanding of what gradients through contact actually mean and where they break down
More latest news
Read all