Contact Modes Are Strata: What Geometric Structure Buys in Discrete-Continuous Planning

arXiv:2608.15541 · cs.RO · Submitted 2026-08-16 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Contact Modes Are Strata".

Dev: Contact-rich manipulation presents a mixed discrete–continuous problem where which contacts are active and how to move while holding them are coupled by a change in dimension.

Rosa: First, who's behind it and why it matters.

Paper summary: Dev: So, looking at "Contact Modes Are Strata: What Geometric Structure Buys in Discrete-Continuous Planning," the authors are essentially showing how geometric structure provides a natural framework for discrete–continuous planning <ref:2608.15541#pg0>.

Rosa: They tackle the core challenge of deciding which contacts are active and how to move while holding them by coupling that choice to a change in dimension <ref:2608.15541#pg0>.

Taro: The title itself, "Contact Modes Are Strata," really captures the main idea: treating the modes not as labels but as actual geometric regions within the configuration space <ref:2608.15541#pg2>.

Dev: It means a plan is fundamentally a walk over these strata, and the discrete mode emerges because we are moving from one stratum to another when we make or break a contact <ref:2608.15541#pg2>.

Rosa: The implication for field robotics is that if we can formalize motion this way, it could allow robots to handle complex manipulation tasks with less explicit programming about every single contact sequence <ref:2608.15541#pg0>.

Taro: If the stratification dictates the gait for a complex task like rotating a cube, that suggests a level of emergent behavior that is very valuable when dealing with unstructured environments <ref:2608.15541#pg0>.

Dev: The preliminary results on pushing and in-hand reorientation showed success in seconds without any prior mode or sequence input, which supports the idea that this geometric structure guides the search effectively <ref:2608.15541#pg0>.

Rosa: So, we're seeing a system where the planner discovers the necessary contact sequence through sampling and projection onto these strata, rather than having it specified beforehand <ref:2608.15541#pg2>.

Taro: The paper suggests that this geometric approach offers a richer way to model manipulation than treating modes as simple labels because it incorporates the underlying dimensional constraints directly <ref:2608.15541#pg0>.

Conclusion: Rosa: So, we've seen how this paper uses geometric stratification to describe contact modes in discrete-continuous systems.

Dev: Yeah, that stratification idea is what really caught my attention from a control engineering standpoint, especially when thinking about loop rates and latency for real-world application.

Taro: I'm wondering how robust this geometric structure is when the environment throws us unexpected noise or misbehaves during execution.

Rosa: Exactly, Taro; it’s about how that underlying geometry helps guide the planner when things go sideways outside of a clean lab setup.

Dev: Right, and looking at who wrote this paper, I see their background leans heavily into robotics theory and configuration space mapping, which suggests a deep dive into the math behind these strata.

Taro: That background makes sense because if you’re dealing with autonomy research, you need that kind of rigorous mathematical foundation to handle those unpredictable situations we discussed.

Rosa: And the conclusion they draw about contact modes being strata is quite powerful; it moves us away from treating them as simple labels and gives them a tangible geometric meaning.

Dev: That tangible geometry is what matters because it gives us a way to quantify the constraints—the dimension reduction based on active contacts—which helps in designing more efficient motion controllers.

Taro: It seems like this work could really impact how we design autonomous systems that need to switch between grasp strategies seamlessly when faced with an unknown situation.

Rosa: I think the real implication is that planning becomes less about guessing sequences and more about navigating a structured space, which should make complex manipulation much more reliable in unstructured settings.

Space and Terrestrial Autonomous Robotic Systems (STARS) Laboratory · University of Toronto Institute for Aerospace Studies (UTIAS)

cs.RO

Submitted: 2026-08-16

Updated: 2026-10-05

Comments: Spotlight paper at the IROS 2026 Workshop on Geometric-Aware Representations in Robotics, Pittsburgh, USA, October 1, 2026

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 78/100

The gist: Contact-rich manipulation presents a mixed discrete–continuous problem where which contacts are active and how to move while holding them are coupled by a change in dimension.

Key concepts

Configuration Space Stratification: This involves dividing the entire possible robot positions into distinct regions called strata based on which body pairs are in contact. This is done using signed distances between bodies to define collision-free space, ensuring that every configuration belongs to exactly one stratum.
Contact Mode as a Stratum: The core idea is that a specific way the robot holds objects (a 'contact mode') corresponds exactly to one of these geometric strata. This means the system doesn't need pre-defined contact sequences; it discovers which contacts are active during planning by moving through these defined regions.
Stratum Graph: The strata are connected not sequentially but by their boundaries, forming a graph. Edges in this graph represent the motion of closing or breaking a contact while maintaining other constraints. Planning involves searching this graph implicitly using an RRT that tracks both position and the active stratum.
Hybrid Path ($\sigma$): A solution is modeled as a hybrid path, which is a sequence of movements where each step lies within a specific stratum ($SA_k$). Consecutive steps connect at points where contacts are made or broken, allowing the planner to transition between different contact modes naturally.

Terminology

Core — 3× in this paper · explained in 2 episodes
IEEE International Conference on Robotics and Automation — 1× in this paper
Trace — 1× in this paper · explained in 2 episodes
University of Toronto Institute for Aerospace Studies — 1× in this paper

Summary

Contact-rich manipulation presents a mixed discrete–continuous problem where which contacts are active and how to move while holding them are coupled by a change in dimension. This paper demonstrates that contact modes are not merely analogous to strata of the configuration space, but rather that a contact mode is one stratum itself, allowing discrete choice to emerge naturally from the planning process.

The Gist

A plan is a walk over strata whose within-stratum segments are geodesics.

How it works: Defining Strata and Configuration Space Stratification

The core of the approach lies in defining stratified configuration spaces for discrete–continuous systems. This involves using signed distances, denoted as φi(q), between body pairs to define the collision-free set Qfree, where φi(q) ≥ 0 for all pairs i. The set of active contacts at any configuration q is defined as A(q) = the set of indices i for which φi(q) = 0. For a fixed subset of contacts A ⊆ I, the configurations where exactly those contacts hold form the set SA, defined by:

SA = q φi(q) = 0 ∀i ∈ A, φj (q) > 0 ∀j /∈ A. Every configuration has one active set, meaning these sets SA are disjoint and cover Qfree. The dimension of each stratum SA is determined by the rank-nullity theorem: dim SA = n − A, where n is the total dimension of the configuration space and A is the number of active contacts.

How it works: Constructing the Stratum Graph

The strata are not laid out sequentially but are glued along their boundaries, forming a stratified space according to Definition 1. A stratum SB meeting cl(SA) holds every contact of A, and its dimension is B − A. The frontier relation in (3), cl(SA) ⊆ G B ≥ A SB, makes the strata the vertices of a graph called the stratum graph. This graph is created by motion rather than enumeration; an edge is made by closing a gap φj while holding hA at zero, and breaking it by opening that gap again. The planner searches this stratum graph implicitly with a sampling-based planner, growing a single rapidly-exploring random tree (RRT) whose nodes carry the configuration and the stratum identified by its active set.

How it works: Planning over Strata

A solution is formulated as a hybrid path σ, which is defined as a walk A0 → A1 → · · · → Am over strata in which the k-th step is a path γk lying in SAk. Consecutive steps meet at a configuration where one contact is made or broken (Figure 2(c)). Since strata of positive codimension have measure zero, the planner samples a set of contacts A together with an ambient configuration and projects that configuration onto the constraint manifold. The resulting projection may land on a stratum other than the sampled one, which is then relabeled by its active constraints.

How it works: Handling Contact Cones and Foliation

The robot's movement is further constrained by considering contact cones and foliation. For an object being manipulated, passive coordinates are introduced as constraint rows if minimum-contact conditions are not met, confining motion to a subset of SA with dimension n−A−np. Furthermore, the contact cone restricts the velocity of the contact point on an object to lie within a cone about the inward contact normal. Holding a contact fixed on the bodies restricts motion differently; making it records that point in the body frame of the object, replacing one row φi = 0 with three rows that pin it. This creates leaves foliation SA into leaves, where motion within a leaf is sticking while motion across leaves is sliding, and each sticking contact removes two further dimensions.

Preliminary Results

The approach was evaluated on two tasks: pushing a T-shaped block around obstacles (n=10) and reorienting a cube in hand (n=19). The planner solved both tasks within seconds with no mode, contact sequence, or stratum given in advance. For the T-shaped block task, the median planning time was 760 ms. For the in-hand cube task, the median planning time was 9.9 s. The results show that for complex tasks like turning a cube by 120°, a gait emerges naturally because a single grasp cannot achieve it, and the stratification dictates this gait rather than being specified beforehand. The contacts made and broken emerge from the search itself rather than from an explicit specification of modes or sequences.

Limitations and Open Problems

The current formulation treats a contact mode as a stratum, which is richer than a simple mode label.

Improvements for AI systems

Here are the specific improvements that can be made to existing AI systems based on this scientific paper, along with what these improved systems could achieve:

The core improvement is a paradigm shift in planning from explicitly defining discrete modes or contact sequences to implicitly discovering them through the geometry of the configuration space.
An AI system can be redesigned to perform Stratified Configuration Space Planning. Instead of relying on traditional mixed discrete-continuous planners that require pre-defined mode families or symbolic domains, this new system would use a collision engine's signed distances and gradients to dynamically define the configuration space structure (the strata).
The improved system will be able to solve complex contact-rich manipulation tasks—such as pushing a T-shaped block around obstacles or dexterous in-hand reorientation of objects—without needing prior knowledge of the exact sequence of contacts or modes.
The system will generate hybrid paths that are intrinsically defined by the constraints:

@ The AI can plan continuous motions (geodesics) while simultaneously determining when to transition between discrete contact states (making or breaking contacts). This allows for smoother, more naturally constrained motion compared to systems that treat contact changes as abrupt mode switches.

The improved system will inherently handle the coupling by a change of dimension in motion planning. When a robot makes a contact, the planner automatically recognizes that its available degrees of freedom are reduced, and it plans within the resulting lower-dimensional manifold defined by the active contacts.
The system will incorporate task-specific constraints (like minimum contact forces or object passivity) directly into the search process to ensure physical feasibility. This means it won't waste time searching in configuration spaces where a robot cannot physically support an object or achieve a desired motion.
The planning search will be guided by an implicit stratum graph rather than requiring its explicit construction beforehand. The planner will sample contacts and configurations, and the resulting active set (the stratum) dictates the next feasible subset of motion, effectively learning the necessary discrete structure during execution/search.
The AI can achieve faster planning times for contact-rich problems by avoiding exhaustive searches over predefined mode sequences. Instead, it uses a sampling-based planner (like RRT) that intelligently samples configurations and immediately identifies the relevant stratum without needing to enumerate all possible modes or contact sequences beforehand.