Proprioceptive Safe Active Navigation and Exploration for Planetary Environments

This paper introduces PSANE, a framework that enables legged robots to safely navigate and explore unknown deformable granular terrains by leveraging proprioceptive leg-terrain interactions to learn traversability models and optimize frontier selection for real-time, goal-directed navigation.

The Gist

Imagine you are trying to walk across a vast, unknown beach at night. You can see the sand, but you can't tell if the ground beneath your feet is firm, or if it's a deep, sticky quicksand that will trap you. If you step in the wrong spot, you might sink and get stuck forever.

This is the exact problem robots face when exploring other planets like Mars. The ground there isn't just rock; it's often loose, sandy soil (regolith) that can swallow a robot whole. Traditional robots rely on "eyes" (cameras and lasers) to see the terrain. But on a planet, looking at the sand doesn't tell you if it's safe to walk on it. A patch of sand might look flat and hard, but underneath, it could be loose powder.

This paper introduces a new robot brain called PSANE (Proprioceptive Safe Active Navigation and Exploration). Instead of just looking, PSANE teaches the robot to "feel" its way through the unknown, using a clever mix of guessing, learning, and careful planning.

Here is how it works, broken down into simple concepts:

1. The "Blindfolded Hiker" Analogy

Imagine you are hiking in a dark forest with a blindfold on. You can't see the path, but you have a very sensitive walking stick. Every time you tap the ground, the stick tells you how hard or soft the soil is.

• Old Way: The robot tries to guess the path based on a blurry map it took from space. It walks straight toward the goal and often falls into a trap.
• PSANE Way: The robot takes a step, feels the ground, and immediately updates its mental map. If the ground feels squishy, it knows, "Okay, I can't go there." If it feels firm, it marks that spot as safe.

2. The "Uncertainty Cloud" (The Magic of Guessing)

The robot can't feel the entire planet at once; it only feels the spots where its feet touch. So, how does it know what's between its steps?

PSANE uses a mathematical tool called a Gaussian Process. Think of this as a "Confidence Cloud."

• When the robot feels a spot, it draws a circle of safety around it.
• Inside the circle, it's very confident the ground is safe.
• As you move away from the footprints, the "cloud" gets fuzzier. The robot knows it's probably safe, but it's not 100% sure.
• The Safety Rule: PSANE is extremely cautious. It only walks where the "cloud" says, "I am 99% sure this is safe." If the cloud is too fuzzy, the robot treats it like a wall and won't go there.

3. The "Explorer's Dilemma" (To Go or To Explore?)

The robot has two conflicting goals:

1. Get to the Goal: "I need to reach that yellow star over there!"
2. Expand the Map: "I need to walk around to find out if there are safe paths nearby."

If it only chases the goal, it might get stuck. If it only explores, it never finishes the mission.

PSANE solves this with a Smart Balancing Act. It looks at the edge of its known safe area (called the "Frontier") and asks two questions for every possible next step:

• Question A: "If I go there, how much new safe ground will I discover?" (Expansion)
• Question B: "How close does this get me to my final goal?" (Progress)

It doesn't just pick the closest spot to the goal. It picks the spot that offers the best deal: a good chance of finding new safe ground while moving forward. It's like a hiker who takes a slight detour to find a bridge, knowing that bridge will save them hours of walking later.

4. The "Reactive Reflex" (The Dance)

Once the robot picks a target spot, it doesn't just drive in a straight line. The ground might change, or a rock might appear.

• PSANE uses a Reactive Controller. Think of this as the robot's reflexes.
• Even while walking toward its target, the robot is constantly dodging "unsafe zones" (areas it hasn't proven safe yet). It's like dancing: you have a destination, but you are constantly adjusting your steps to avoid tripping over invisible obstacles.

Why is this a Big Deal?

In the past, robots on Mars (like the Spirit rover) got stuck because they looked at the ground, thought it was safe, and then sank. They relied too much on "eyes" and not enough on "feel."

PSANE changes the game by:

• Trusting the feet over the eyes: It uses the physical sensation of walking to build a map.
• Being paranoid (in a good way): It assumes the unknown is dangerous until proven otherwise.
• Planning smartly: It balances the urge to explore with the need to finish the job.

The Bottom Line

This paper presents a robot that is brave enough to explore the unknown, but smart enough to know when to stop. It turns the scary problem of "walking on unknown, sinking sand" into a solvable puzzle by constantly feeling the ground, updating its map, and taking calculated steps toward the goal. It's the difference between a robot that gets stuck in the sand and one that successfully walks across the dunes.

Technical Summary

Here is a detailed technical summary of the paper "Proprioceptive Safe Active Navigation and Exploration for Planetary Environments".

1. Problem Statement

The paper addresses the challenge of navigating legged robots in unknown, deformable granular terrains (e.g., lunar regolith or Martian soil) where traditional exteroceptive sensing (vision, LiDAR) is insufficient.

• The Limitation: In deformable environments, surface geometry does not reliably predict traversal risks. Critical factors like shear strength, compaction, and sinkage resistance are subsurface properties that cannot be inferred remotely. Relying solely on visual data can lead to immobilization (e.g., the Spirit rover incident).
• The Core Challenge: The robot must navigate to a specific goal while ensuring safety, using only proprioceptive data (leg-ground interaction forces and depths). This creates a coupled problem of:
  1. Estimation: Inferring a continuous, spatially varying risk map (slip ratio) from sparse, noisy, discrete contact measurements.
  2. Exploration: Actively acquiring data to expand the "certified safe" region.
  3. Control: Executing a trajectory that balances reaching the goal with expanding the safe set, all while avoiding uncertified (potentially unsafe) regions.

2. Methodology: The PSANE Framework

The authors propose PSANE (Proprioceptive Safe Active Navigation and Exploration), a framework that integrates online learning, safety certification, and reactive control.

A. Proprioceptive Terrain Mapping (Gaussian Processes)

• Data Source: The robot estimates local terrain penetration resistance and slip ratio ($f(x)$) at discrete contact points using a "rotary-walking" model based on force-depth measurements.
• Modeling: A Gaussian Process (GP) regression is used to learn the slip-ratio field $f: \mathcal{W} \to \mathbb{R}$.
  • The GP provides a predictive mean $\mu_t(x)$ and variance $\sigma_t^2(x)$ for the slip ratio at any location.
  • This allows the robot to infer terrain properties in unvisited areas based on sparse interactions.

B. Uncertainty-Aware Safe-Set Certification

To ensure safety, the system does not rely on the mean prediction alone but uses confidence bounds to certify regions.

• Confidence Interval: A conservative upper bound is defined as $u_t(x) = \mu_{t-1}(x) + \sqrt{\beta}\sigma_{t-1}(x)$, where $\beta$ controls the conservatism.
• Lipschitz Continuity: Assuming the terrain properties vary smoothly (Lipschitz constant $L$), safety is propagated from known safe points to neighbors. A point $x'$ is certified safe if:
  $$u_t(x) + L\|x - x'\| \leq h$$
  where $h$ is the safety threshold (maximum allowable slip).
• Result: This generates a conservative "Safe Region" ($S_t$) that expands iteratively as uncertainty decreases.

C. Multi-Objective Frontier Subgoal Selection

The robot identifies frontiers (boundaries between the safe region and unknown terrain) as candidates for the next subgoal. Selection is formulated as a multi-objective optimization balancing:

1. Safe-Set Expansion Probability ($P_e$): The likelihood that moving to a frontier will allow the robot to certify new safe areas. Calculated based on the number of uncertified states that can be guaranteed safe from that frontier.
2. Goal-Directed Cost ($V_{goal}$): A heuristic value based on the distance to the mission goal and the current robot position.

Selection Strategy:
The authors propose a scalarization approach (PSANE) over the Pareto-optimal frontier set:
$$V_{overall}(x) = V_{goal}(x) \cdot P_e(x)$$
This single score allows the robot to dynamically choose subgoals that maximize progress toward the goal while simultaneously expanding the safe map, avoiding the inefficiency of alternating strictly between "explore" and "go" phases.

D. Reactive Navigation

Once a subgoal is selected, a diffeomorphic reactive controller (based on potential fields) generates real-time velocity commands.

• The certified safe region is converted into a geometric free-space polygon.
• Uncertified regions are treated as obstacles with a safety margin.
• The controller ensures smooth, obstacle-free trajectories without requiring frequent global replanning.

3. Key Contributions

1. Problem Formulation: A novel formulation of safe navigation on deformable terrain as a continuous risk estimation problem from sparse proprioceptive data, incorporating uncertainty-aware safety certification.
2. Algorithm Design: A multi-objective frontier selection strategy that balances safe-set expansion and goal progress via scalarization, integrated with a reactive controller for real-time execution.
3. Validation: Extensive simulation using field-measured LHS-1 regolith terramechanics data, demonstrating robust performance in both smooth and highly heterogeneous terrain environments.

4. Experimental Results

The framework was tested in a Chrono simulation using LHS-1 lunar simulant parameters, comparing PSANE against three baselines:

• NGH (Naive Goal Heuristic): Direct path to goal.
• SGH (Safe-Goal Heuristic): Safety-aware but no active expansion.
• PGH (Pareto-Goal Heuristic): Alternates between objectives without scalarization.

Key Findings:

• Success Rate:
  • NGH: 0% success (robots immobilized in high-slip zones).
  • SGH: 66.7% (Env 1) and 0% (Env 2); frequently stalled at safe-region boundaries.
  • PSANE & PGH: 100% success in both environments.
• Efficiency:
  • PSANE significantly outperformed PGH in efficiency. In Environment 1, PSANE reduced completion time by ~46% (111s vs 209s) and path length by ~54% (14.5m vs 31.9m).
  • This is attributed to the scalarized objective, which avoids the "detours" caused by the rigid phase-switching of the PGH baseline.
• Exploration: In goal-free exploration tasks, PSANE achieved higher coverage (0.67 vs 0.59) over 1000 seconds due to its cost-aware frontier selection.

5. Significance

This work bridges the gap between locomotion control and high-level navigation for planetary exploration.

• Safety-Critical Autonomy: It provides a rigorous mathematical framework for certifying safety in environments where remote sensing fails, crucial for future missions to the Moon or Mars.
• Active Learning: It demonstrates how robots can actively learn terrain properties to improve their own navigation capabilities in real-time.
• Scalability: The approach is designed for single-robot deployment but lays the groundwork for multi-robot coordination and longer-horizon decision-making in uncertain, deformable environments.