Imagine walking through your house in pitch darkness. You do not freeze; instead, your entire locomotive strategy shifts. You extend your feet cautiously, feeling for the edge of a rug or the rise of a step. When your toe bumps a chair leg, you do not need to see it to know it is there—your joints feel the resistance, and your brain instantly triggers a reflex to step over it.
In robotics, this internal sense of body position, force, and motion is called proprioception. While computer vision and LiDAR have long been the primary tools for robot navigation, they frequently fail in low-visibility environments like smoke, darkness, or dust, and they can be fooled by transparent glass or soft terrain like tall grass.
A powerful research direction has emerged to solve this vulnerability: enabling robots to walk, climb, and navigate complex terrains "blindly" by treating their own internal physical dynamics as a rich, real-time mirror of the external world.
The Physics of the Internal Mirror
Why can a robot understand the world without seeing it? The answer lies in classical mechanics: every action has an equal and opposite reaction. When a robot interacts with its environment, the terrain exerts forces back onto the robot's body. These forces propagate through the limbs and manifest as measurable changes in internal states, such as motor currents, joint velocities, and body orientation.
Early work in this domain relied on explicit physical models to capture this relationship. For instance, Schröder-Schetelig et al. utilized the biological concept of an efference copy—a duplicate of a commanded motor signal—alongside a forward internal model. By comparing what a motor was commanded to do with what it actually did, the system could immediately deduce changes in terrain slope.
[Commanded Action] ───> [Forward Model] ───> (Expected State)
│
[Compare Discrepancy] ───> Infer Terrain/Obstacle
│
[Actual Motor Action] ─────────────────────> (Actual State)
This discrepancy-based intuition applies directly to obstacle detection. If a balancing robot or a manipulator arm strikes an object, its motors must draw more current to fight the obstruction. Searock et al. demonstrated that monitoring the gap between commanded and actual velocity, alongside motor current spikes, provides sufficient data to reliably trigger safety reflexes during collisions. Similarly, Fukushima et al. showed that multi-degree-of-freedom manipulators can identify physical contact autonomously by evaluating motor torque data against the robot's internal physics model.
Even unconventional geometries leverage this principle: Wang et al. showed that snake robots can infer environmental obstructions purely by monitoring deviations in their body wave amplitudes, allowing them to dynamically adjust their compliance to push past or yield to obstacles.
From Instantaneous Reactions to Temporal History
While a single snapshot of a joint torque spike can signal an immediate collision, it is highly ambiguous. Is the robot stepping on a hard rock, sinking into mud, or experiencing a brief motor glitch?
To understand the type of terrain it is traversing, a robot needs context. Modern reinforcement learning (RL) frameworks achieve this by analyzing a continuous temporal history of proprioceptive data. As a robot moves across a surface, the sequence of its joint positions, velocities, and errors forms a distinct physical signature.
In the seminal Rapid Motor Adaptation (RMA) framework, Kumar et al. demonstrated that a short history of joint states and recent actions implicitly encodes complex environmental parameters like friction, payload, and ground compliance. Instead of manually engineering physics equations for every possible surface, an adaptation module learns to map this movement history to an environmental vector in real time, allowing a quadruped to transition from dry ground to mud or sand in fractions of a second.
Taking this further, Long et al. framed environmental factors like terrain height and friction as external "disturbances." By training a network to analyze how the robot's inertial measurement unit (IMU) and joint encoders react to these disturbances, the system uses the robot's physical response as a direct proxy for external data. This allows frameworks like DreamWaQ, developed by Nahrendra et al., to navigate unstructured environments by using an estimator network to "imagine" the underlying terrain properties purely from internal feedback.
Reconstructing the World Through Touch
If a temporal history of proprioception can encode abstract properties like friction, it can also encode explicit geometry. When a robot's leg collides with a step, the exact angle of the leg and the height at which the foot stops contain precise geometric information about that step.
This insight allows blind robots to execute complex maneuvers that previously required vision. For example, Chamorro et al. trained legged and wheeled-legged robots to climb stairs blindly. By treating physical resistance from joint and wheel encoders as a proxy for stair geometry, the control policy automatically triggers specialized obstacle-clearing reflexes—such as lifting a leg higher—without any visual map.
On a larger scale, robots can use their limbs like a cane to map their surroundings:
- Height Mapping: Gu et al. introduced a denoising world model that enables humanoid robots to reconstruct local terrain height maps purely from joint torques and inertial data, ensuring stable locomotion over irregular surfaces and snow.
- Clutter Estimation: The PROBE framework by Ramesh et al. demonstrates that a legged robot can reconstruct both the geometry and the mobility (whether an object is rigid or pushable) of its surroundings by analyzing joint torques and body pose variations as it moves through clutter.
Overcoming the Lack of Foresight
The fundamental limitation of blind locomotion is the absence of a visual preview. A blind robot cannot prepare for a drop-off or a hurdle before reaching it. As Apostolides et al. note, performing explosive maneuvers like jumping without a visual preview requires the control policy to be exceptionally resilient, relying entirely on real-time IMU and joint feedback to catch and stabilize the body during high-impact landings.
To navigate complex spaces safely without foresight, a robot must substitute a forward-looking camera with a backward-looking memory. If a robot bumps into an obstacle, it must remember that obstacle's location so it does not repeatedly try to walk through it.
[Blind Movement] ───> [Physical Collision] ───> [Store Contact Keypoint in Memory]
│
[Path Replanning] <─── [Condition Policy on Memory] <──────────┘
To implement this, Bejarano et al. developed ProDapt, a system that conditions diffusion-based control policies on a curated memory of historical contact points. By explicitly remembering past physical collisions, the robot avoids getting trapped in infinite loops when navigating blind, cluttered spaces.
This tactile memory can also complement exteroceptive sensors when they are available. In the VP-Nav system detailed in Coupling Vision and Proprioception for Navigation of Legged Robots, when a robot physically encounters an obstacle invisible to its cameras (such as a glass wall), the proprioceptive contact feedback dynamically inserts a virtual obstacle into the global occupancy map, forcing the global planner to reroute.
Ultimately, proprioception provides the foundational safety net for robotics. When visual-inertial odometry drifts or fails entirely due to darkness or smoke, preintegrating high-frequency leg kinematics and contact factors—as demonstrated by Hartley et al.—provides a fail-safe localization tracking method that keeps the robot oriented, stable, and moving forward.
Key Foundational Work
Learning to Prevent Failure States for a Dynamically Balancing Robot (2005) — contributed 2 concept elements to this direction.
1P1-F4 Development of the New Motor Driver and Introduction of the Products in Collaboration (2001) — contributed 2 concept elements to this direction.