Contact-Triggered Blind Climbing (CTBC)

• CTBC is a control paradigm where tactile and proprioceptive contact events trigger adaptive climbing without relying on external vision or mapping.
• It is implemented across various morphologies—legged, wheeled, many-legged, and limbless—each reconfiguring motion control upon local contact detection.
• The approach enables robust stair ascent and vertical obstacle negotiation, validated through simulations and hardware experiments across diverse terrains.

Searching arXiv for the cited CTBC-related papers to ground the article. Contact-Triggered Blind Climbing (CTBC) is a climbing regime in which tactile or proprioceptive contact events, rather than vision, mapping, or localization, trigger the control transitions required for stair ascent, obstacle negotiation, or vertical climbing. In the literature summarized here, CTBC appears in several embodiments: blind stair ascent for legged and wheeled-legged robots using proprioception and learned reflexes; tactile obstacle climbing for elongate many-legged robots; contact-triggered leg-lifting for wheeled bipeds; and whole-body force redistribution in cornsnakes climbing a smooth vertical wall with discrete posts. Across these systems, the common structure is event-driven adaptation after contact formation, with deployment operating without exteroceptive sensing (Chamorro et al., 2024, He et al., 11 Apr 2025, Li et al., 3 Sep 2025, Riiska et al., 6 Jul 2026).

1. Conceptual definition and scope

CTBC denotes a control principle in which the decisive signal for climbing is the occurrence of contact itself. In the wheeled-biped formulation, CTBC “solves obstacle and stair traversal for wheeled bipedal robots without exteroceptive sensing,” and the trigger is the moment when wheel-obstacle contact exceeds a threshold in the horizontal-plane contact force (Li et al., 3 Sep 2025). In the many-legged formulation, all climbing decisions and control actions are triggered exclusively by tactile sensing from a whisker-like antenna and binary foot contact sensors; no visual sensing is used or required (He et al., 11 Apr 2025). In the blind stair-climbing reinforcement-learning formulation, the controller relies only on proprioception and contact signatures, plus an externally provided terrain boolean, and the policy learns to “feel” an obstacle through abrupt deceleration or joint and drive dynamics changes when a wheel or foot hits a riser (Chamorro et al., 2024). In the limbless vertical-climbing formulation, CTBC is the regime in which a limbless body ascends or descends sparsely featured vertical surfaces by using tactile contact events to trigger a fast, stereotyped, whole-body redistribution of forces that preserves global equilibrium while incorporating new footholds (Riiska et al., 6 Jul 2026).

The term “blind” does not imply sensorless operation. In all four cases, control depends on onboard signals: IMU measurements, joint encoders, contact forces, foot contacts, or tactile antenna readings. What is excluded is exteroceptive perception such as vision, terrain reconstruction, SLAM, GPS, or localization at deployment (Chamorro et al., 2024, He et al., 11 Apr 2025, Li et al., 3 Sep 2025). In this sense, CTBC is best understood as contact-driven locomotor adaptation under partial observability.

A further distinction concerns the scale of the response. In wheeled bipeds and legged stair climbing, contact triggers a phase update or a stepping reflex. In limbless climbing, contact triggers a system-wide redistribution of force within a redundant contact network. This suggests that CTBC is not tied to a single morphology; rather, it describes a class of controllers whose primary event variable is contact formation.

2. Sensory basis and event triggering

The sensory substrate of CTBC is local and low bandwidth. In the 2024 blind stair-climbing controller, the actor observation consists of IMU angular velocity $\omega_t \in \mathbb{R}^3$, projected gravity $\gamma_t \in \mathbb{R}^3$, planar goal direction $g_{\text{dir}} \in \mathbb{R}^2$, heading error $\theta_{\text{err}}$, a base-height command $h_{\text{target}}$, a terrain boolean $b \in \{0,1\}$, joint positions and velocities, and the last action. The actor does not receive contact forces at runtime; instead, it infers riser contact from proprioceptive signatures such as a sudden decrease in base forward velocity, wheel or foot deceleration, or joint deflection (Chamorro et al., 2024). In the Tron1 framework, the trigger is explicit: if the magnitude of the contact force in the $xy$ plane for foot $i$ exceeds a threshold, then a leg-lift motion is initiated. To suppress noise, a three-frame sliding window is applied, and stable contact is declared only if all three recent frames exceed threshold (Li et al., 3 Sep 2025).

The many-legged robot uses two tactile channels. The head-mounted antenna is coupled via two springs to two Force Sensing Resistors with analog output $s_a(t)\in[0,1024]$, with “Hit” defined by $s_a(t)>\tau_a$ and $\gamma_t \in \mathbb{R}^3$0; the antenna oscillates vertically as $\gamma_t \in \mathbb{R}^3$1 with $\gamma_t \in \mathbb{R}^3$2 and $\gamma_t \in \mathbb{R}^3$3. Toe-tip capacitive sensors provide analog foot signals $\gamma_t \in \mathbb{R}^3$4, with contact when $\gamma_t \in \mathbb{R}^3$5 and $\gamma_t \in \mathbb{R}^3$6. These foot contacts are aggregated into segmental duty factors $\gamma_t \in \mathbb{R}^3$7, and “floating segments” are defined by $\gamma_t \in \mathbb{R}^3$8 with $\gamma_t \in \mathbb{R}^3$9 (He et al., 11 Apr 2025). In the cornsnake study, the trigger is the engagement of a new post contact, detected experimentally through per-post load cells and kinematics, after which force on that post and its neighbors ramps to a characteristic, stable level (Riiska et al., 6 Jul 2026).

Embodiment	Primary trigger signal	Immediate control response
Blind stair climbing	Proprioceptive signature of riser contact	Raise contacting limb, shift body, step trailing limb
Many-legged robot	Antenna hit/void and foot-contact duty factor	Head raise or pitch-down; edge press near floating segment
Wheeled biped (Tron1)	Stable $g_{\text{dir}} \in \mathbb{R}^2$0-plane contact force over a three-frame window	Trigger feedforward leg lift
Cornsnake vertical climbing	New foothold engagement in a redundant contact network	Whole-body force redistribution

Because the trigger is local, CTBC generally dispenses with explicit terrain geometry at runtime. The many-legged system reconstructs only short-range contour proxies through the antenna-tip height history, using $g_{\text{dir}} \in \mathbb{R}^2$1 and $g_{\text{dir}} \in \mathbb{R}^2$2 computed from recent hits (He et al., 11 Apr 2025). The stair-climbing RL systems do not construct a terrain map for the actor at all (Chamorro et al., 2024, Li et al., 3 Sep 2025). The snake study is even more direct: the wall is smooth except for posts, and contact detection itself is the basis for reorganization of the force network (Riiska et al., 6 Jul 2026).

3. Control laws and embodiment-specific implementations

In legged and wheeled-legged blind stair climbing, the central design choice is position-based reinforcement learning. The policy outputs articulated-joint position targets and wheeled-joint angular velocities, which are tracked by low-level PD controllers. The study argues that position-based action parameterization is vital because stairs require phase-dependent motion with non-linear profiles: approach acceleration, deceleration on riser contact, step-up, and re-acceleration. Velocity-tracking controllers, by contrast, constrain the robot to follow a preset speed and struggle to discover or stabilize the needed sequence (Chamorro et al., 2024). A stair-mode boolean $g_{\text{dir}} \in \mathbb{R}^2$3 yields two specialized behaviors within one policy: with $g_{\text{dir}} \in \mathbb{R}^2$4, slower gait, larger leg split, higher swing clearance, and contact-triggered stepping; with $g_{\text{dir}} \in \mathbb{R}^2$5, faster gait and standard locomotion on non-stair terrain (Chamorro et al., 2024).

The many-legged controller is a hybrid tactile feedback law superimposed on open-loop body and leg waves. The antenna tip height relative to the head joint is

$g_{\text{dir}} \in \mathbb{R}^2$6

From recent hit-associated heights, the controller computes $g_{\text{dir}} \in \mathbb{R}^2$7 and $g_{\text{dir}} \in \mathbb{R}^2$8 and uses a contact-triggered head controller:

$g_{\text{dir}} \in \mathbb{R}^2$9

The gains are $\theta_{\text{err}}$0, $\theta_{\text{err}}$1, the offset is $\theta_{\text{err}}$2, and the descent bias is $\theta_{\text{err}}$3. A second controller uses foot-contact duty factors to detect floating segments and commands the vertical joint immediately ahead of the smallest-index floating segment as $\theta_{\text{err}}$4, with $\theta_{\text{err}}$5 and $\theta_{\text{err}}$6 (He et al., 11 Apr 2025). The result is a local “raise/engage/drag/descent” state machine driven entirely by contact events.

In Tron1, CTBC combines instruction learning and reinforcement learning. The trigger condition is stable contact force at a wheel or foot, and the triggered motion is a strongly-guided feedforward trajectory injected into the hip-pitch and knee-pitch joints, with the knee trajectory amplitude set to twice the hip amplitude. The composite action is

$\theta_{\text{err}}$7

with a half-cosine feedforward profile

$\theta_{\text{err}}$8

As training progresses, $\theta_{\text{err}}$9 is annealed to zero, so deployment uses the learned network policy alone (Li et al., 3 Sep 2025).

The limbless case differs in implementation but not in logic. There, CTBC is the incorporation of a new contact by a balance-preserving whole-body force update. The paper proposes a robotic implementation in which a small quadratic program selects a new force distribution within the nullspace of the balance constraints while satisfying friction cones and effort costs. That optimization is explicitly described as not being used in the biological experiments but as consistent with the measured redistribution mechanism (Riiska et al., 6 Jul 2026).

4. Mechanics, balance, and force redistribution

A defining mechanical feature of CTBC in the limbless study is redundant contact. Cornsnakes on closely spaced posts maintained 8–16 simultaneous contacts, and on widely spaced posts 5–7 contacts, far exceeding the three contacts needed to satisfy quasi-static planar equilibrium constraints (Riiska et al., 6 Jul 2026). Whole-body balance is expressed as

$h_{\text{target}}$0

With stacked per-post forces $h_{\text{target}}$1 and a balance matrix $h_{\text{target}}$2, the constraint set is $h_{\text{target}}$3. When a new contact forms, the rearrangement $h_{\text{target}}$4 decomposes into a balance-preserving nullspace component and a small residual outside the nullspace. The magnitude of the nullspace redistribution obeys

$h_{\text{target}}$5

with $h_{\text{target}}$6 across conditions (Riiska et al., 6 Jul 2026). This is the most explicit mechanical formulation of CTBC in the provided literature: contact triggers a rapid update in internal force allocation while net wrench remains approximately unchanged.

The snake study also shows that passive mechanics can be sufficient, but not exhaustive, for CTBC. A quasi-static computational model and a physical robot both climb using only passive Coulomb friction at contacts and a prescribed body wave. However, ascending snakes depart systematically from this passive baseline. For each dynamic contact, tangential power is defined as $h_{\text{target}}$7; during descent, most contacts are dissipative with $h_{\text{target}}$8 and effective kinetic friction $h_{\text{target}}$9, whereas during ascent more than 40% of dynamic contacts are active with $b \in \{0,1\}$0 across conditions (Riiska et al., 6 Jul 2026). A common misconception is therefore that blind contact-based climbing is necessarily passive. The limbless results contradict that interpretation: contact-triggered control can include active power injection at selected contacts.

In the legged and wheeled-legged stair-climbing formulations, the relevant mechanical principle is phase-dependent clearance under uncertain terrain geometry. The 2024 study argues that position targets allow the policy to discover approach deceleration on impact, high-clearance step-up, and re-acceleration, while the critic’s privileged access to terrain and contact signals shapes the value function during training (Chamorro et al., 2024). Tron1 uses the same contact-to-clearance principle in a more explicit state machine: stable horizontal-plane contact triggers a leg-lift motion because continuous rolling alone cannot generate sufficient vertical lift for steps near or above the wheel radius (Li et al., 3 Sep 2025).

The many-legged tactile controller relies less on precise force regulation and more on redundancy and static stability. The robot has high static stability and many simultaneous contacts, so binary hit/void sensing and foot-contact duty factors are sufficient to coordinate a small number of joint overrides. This suggests that CTBC can trade sensing richness against morphological redundancy: the limbless case uses detailed force redistribution over many contacts, whereas the many-legged case uses sparse tactile events to modulate a highly stable gait (He et al., 11 Apr 2025, Riiska et al., 6 Jul 2026).

5. Learning, privileged information, and sim-to-real transfer

Two of the robotic CTBC instantiations are learned with reinforcement learning, and both use asymmetric actor-critic designs. In the 2024 blind stair-climbing method, PPO is used with an actor that receives only proprioception and commands, while the critic additionally receives privileged information such as linear velocity, absolute base height, relative goal position, terrain height measurements from a $b \in \{0,1\}$1 grid with $b \in \{0,1\}$2 values, contact forces filtered with a 5-timestep sliding window, and terrain friction $b \in \{0,1\}$3 (Chamorro et al., 2024). In Tron1, the actor uses IMU angular velocity, projected gravity, joint positions and velocities, and last actions, while the critic receives privileged linear velocity and average contact forces. A state-estimation MLP encoder is co-trained as part of the asymmetric actor-critic architecture (Li et al., 3 Sep 2025).

Both works stress that contact-triggered blind climbing is not learned as generic locomotion alone. In the 2024 study, environments are synthesized in a $b \in \{0,1\}$4 curriculum with increasing difficulty across 12 rows, spanning discrete steps, multi-step stair flights, smooth slopes, pyramids, discrete obstacles, and flat terrain (Chamorro et al., 2024). In Tron1, the curriculum is an $b \in \{0,1\}$5 arena split into 10 columns and 10 rows, with smooth and rough slopes, stair columns, and discrete-obstacle columns, and training is conducted in two stages: first without a lateral-foot-position constraint, then with a “same foot x position” reward to correct the fore-aft gait acquired earlier (Li et al., 3 Sep 2025).

Sim-to-real robustness is addressed through domain randomization and delay modeling. In the 2024 stair-climbing work, friction is randomized, random pushes inject base linear velocities up to $b \in \{0,1\}$6 at most every 3 s, and delay randomization simulates approximately 20 ms latency by using the previous state with 50% probability. The study reports that this delay randomization was crucial: without it, real-world deployments exhibited unstable, jerky motion with joint velocities up to $b \in \{0,1\}$7, whereas with it, joint velocities remained under $b \in \{0,1\}$8 and climbs were smooth (Chamorro et al., 2024). Tron1 randomizes payload mass, center-of-mass shift, PD factors, friction, restitution, inertia, motor torque, IMU offset, default joint positions, step delay in $b \in \{0,1\}$9, and external pushes at 8 s intervals with $xy$0 planar velocity (Li et al., 3 Sep 2025).

The role of contact guidance is also explicit in the ablations. In the 2024 study, removing privileged information (“No priv.”) dramatically reduces stair success rates across robots, and removing the terrain boolean compromises the policy’s ability to specialize between stair and non-stair terrain (Chamorro et al., 2024). In Tron1, success degrades substantially when either the feedforward component or the contact trigger is removed, and more severely when both are removed (Li et al., 3 Sep 2025). A plausible implication is that CTBC, at least in current robotic realizations, benefits from both event-triggered structure and a training scaffold that makes the triggered response easy to discover.

6. Empirical performance, limitations, and open questions

The reported performance spans multiple morphologies and terrains. In the 2024 blind stair-climbing study, the method is evaluated on Unitree Go1, Cassie, Ascento, and ANYmal on Wheels. Simulation results report, for example, Ascento at 82–88% success for 14–18 cm steps with stair mode on, Cassie at approximately 100% up to 22 cm, and ANYmal on Wheels in biped mode at approximately 93% across 14–30 cm with stair mode on. In the real world, Ascento repeatedly climbed 15 cm steps, with the learned reflex lifting the contacting side first and then stabilizing on the tread (Chamorro et al., 2024).

The tactile many-legged robot shows a different performance envelope. Open-loop lateral and vertical wave templates alone cross 5–10 cm obstacles and are blocked at 15 cm, whereas the CTBC feedback controller succeeds on box obstacles up to 20 cm and bundled-cylinder obstacles up to 25 cm, which is approximately five times the robot’s center belly height of about 5 cm. Outdoor tests include 20 cm obstacle climbing in mud, grass, boulders, leaves, and confined spaces, and pipe inspection in a 0.2 m radius, 3 m long pipe completed in 80 s, or approximately $xy$1 (He et al., 11 Apr 2025).

Tron1 validates CTBC on a wheeled biped with smaller wheels than Ascento. Simulation ablations over 100 robots show success rates for CTBC of 100% at 8, 10, and 12 cm, 98% at 15 cm, 96% at 18 cm, 86% at 20 cm, and 70% at 22 cm. The corresponding “w/o contact-trigger” condition falls to 62%, 60%, 56%, 46%, 18%, 2%, and 0%, and “w/o both” falls to 46%, 34%, 28%, 8%, 4%, 0%, and 0%. On hardware, Tron1 escapes a 10 cm hole and climbs open-gap stairs of 16 cm and 20 cm, including continuous ascent of three 20 cm steps, while using only a 10 cm feedforward lift reference during training guidance (Li et al., 3 Sep 2025).

The limbless study reports complete success of all cornsnakes climbing up and down in all post configurations, with speed depending on post geometry: long posts support faster lateral undulation, whereas short posts produce slower motion with mixed lateral-undulation and concertina-like behavior. The biological comparison is also informative: cornsnakes succeeded on the posted wall, while kingsnakes failed, indicating that CTBC depends not merely on contact availability but on appropriate kinematic and force modulation (Riiska et al., 6 Jul 2026).

Several limitations recur across embodiments. The 2024 stair-climbing controller relies on an externally toggled terrain boolean, so stair mode is not inferred autonomously (Chamorro et al., 2024). The many-legged system can misinterpret ceilings as climbable obstacles, and its rigid-base antenna can jam; future work in that paper proposes filtering and TPU-printed antenna designs (He et al., 11 Apr 2025). Tron1 does not report a numerical contact threshold and exhibits an early bias toward lifting the front leg first due to the first training stage (Li et al., 3 Sep 2025). In the limbless case, performance depends strongly on feature spacing, post length, and frictional coupling, and shorter or sparser features increase lateral bracing demands and residual balance violations (Riiska et al., 6 Jul 2026).

Taken together, these results define CTBC as a control paradigm rather than a single algorithm. Its invariant features are local contact detection, event-triggered reconfiguration, and operation without exteroceptive perception. Its embodiment-specific differences lie in what is reconfigured: joint targets and gait phase in legged systems, vertical spine posture in many-legged systems, or a nullspace force distribution in limbless systems. The present literature also indicates that “blind” climbing need not be purely reactive in a narrow sense: it can be supported by privileged-information training, feedforward instruction, or redundant-contact mechanics, provided that deployment decisions remain grounded in contact-triggered state changes rather than terrain reconstruction (Chamorro et al., 2024, He et al., 11 Apr 2025, Li et al., 3 Sep 2025, Riiska et al., 6 Jul 2026).