PaperlayerView exact revision on arXiv

Source-linked AI summary

Autonomous Spot: Long-Range Autonomous Exploration of Extreme Environments with Legged Locomotion

Amanda Bouman, Muhammad Fadhil Ginting, Nikhilesh Alatur, Matteo Palieri, David D. Fan, Thomas Touma, Torkom Pailevanian, Sung-Kyun Kim, Kyohei Otsu, Joel Burdick, Ali-akbar Agha-mohammadi

arXiv:2010.09259v3cs.RO

TL;DR

The paper addresses how to achieve large-scale, long-duration autonomous exploration with a legged Spot robot in extreme environments. It integrates NeBula’s risk- and perception-aware autonomy with Spot mobility, sensing, odometry, planning, and hardware, and demonstrates the resulting system in real-world DARPA SubT missions. The system explored, mapped, traversed difficult terrain, and detected artifacts under mission constraints.

  • Problem

    Extreme underground exploration requires legged robots to perform complex, long-duration missions under time, communication, computation, mobility, and perceptual constraints.

  • Method

    The paper integrates NeBula’s risk- and perception-aware autonomy architecture with Boston Dynamics’ Spot, including multi-sensor odometry, perception- and traversability-aware planning, and mission-level autonomy.

  • Results

    The system was field-tested in DARPA SubT Urban Circuit missions, where Au-Spots explored and mapped difficult environments, traversed stairs and rubble, and detected artifacts.

  • Takeaways & Limitations

    The demonstrated integration advances autonomous legged exploration by showing Spot-based NeBula autonomy operating in complex real-world underground missions.

Abstract

from arXiv · show

This paper serves as one of the first efforts to enable large-scale and long-duration autonomy using the Boston Dynamics Spot robot. Motivated by exploring extreme environments, particularly those involved in the DARPA Subterranean Challenge, this paper pushes the boundaries of the state-of-practice in enabling legged robotic systems to accomplish real-world complex missions in relevant scenarios. In particular, we discuss the behaviors and capabilities which emerge from the integration of the autonomy architecture NeBula (Networked Belief-aware Perceptual Autonomy) with next-generation mobility systems. We will discuss the hardware and software challenges, and solutions in mobility, perception, autonomy, and very briefly, wireless networking, as well as lessons learned and future directions. We demonstrate the performance of the proposed solutions on physical systems in real-world scenarios.

I. INTRODUCTION

The paper targets autonomous exploration, mapping, and search in extreme underground environments, using Spot and NeBula to address the mobility, perception, and autonomy demands of complex missions.

  • I. INTRODUCTION: Extreme-environment missions require robots to explore, map, and search underground settings under challenging operational conditions.The DARPA Subterranean Challenge is the paper’s concrete mission context.
  • I. INTRODUCTION: SubT systems must fit through passages as narrow as 80 cm while carrying substantial sensing, autonomy, computing, and communication payloads.They must also remain operational for missions lasting at least one hour across multi-kilometer areas.
  • I. INTRODUCTION: Legged robots are suited to difficult environments because they can address locomotion, size, payload, and endurance requirements that challenge wheeled systems.Prior work had demonstrated autonomous legged exploration in rough or tunnel-like settings, while this paper focuses on a more comprehensive system-level mission.
  • I. INTRODUCTION: The work integrates NeBula autonomy with Boston Dynamics’ Spot to support complex autonomous missions on a legged platform.The paper presents a system-level overview while detailing selected algorithms critical to legged autonomy.
  • I. INTRODUCTION: The proposed advances include high-level autonomy, reliable multi-sensor odometry, perception- and traversability-aware planning, and a rugged lightweight hardware payload.These capabilities were field-tested during the DARPA SubT Urban Circuit and its practice runs.

II. NEBULA AUTONOMY

NeBula is a probabilistic, risk- and perception-aware autonomy architecture for uncertain environments, coordinating sensing, belief maintenance, and planning to improve world-model accuracy.

  • II. NEBULA AUTONOMY: NeBula was developed for autonomous exploration of extreme planetary and subsurface terrains, including unknown and hard-to-access environments.Its methods are intended to support reliable, coordinated multi-robot exploration.
  • II. NEBULA AUTONOMY: The architecture probabilistically fuses sensing information while predicting and assessing outcomes and risks in uncertain settings.This allows uncertainty to be incorporated into autonomy decisions.
  • II. NEBULA AUTONOMY: NeBula connects Spot interfaces, odometry, belief management, mission planning, and other planning modules into a high-level autonomy architecture.The belief manager constructs and maintains the robot’s model of the environment.
  • II. NEBULA AUTONOMY: Its belief prediction module enables perception-aware and uncertainty-aware planning by improving the accuracy of the world representation and reducing risk.The planner accounts for perceptual capability when selecting actions.

III. AU-SPOT MOBILITY SYSTEM

Au-Spot combines Spot’s quadrupedal mobility with an expanded sensing package and protected auxiliary power and computing hardware for challenging autonomous missions.

  • III. AU-SPOT MOBILITY SYSTEM: Spot is a quadrupedal robot designed to negotiate challenging terrain, including steps and stairs that traditional wheeled robots may not traverse.This mobility platform forms the basis of Au-Spot.
  • III. AU-SPOT MOBILITY SYSTEM: Au-Spot denotes Spot powered by the NeBula autonomy package.The system combines the robot, NeBula sensors, and auxiliary payload hardware.
  • III. AU-SPOT MOBILITY SYSTEM: NeBula augments Spot’s factory perception with LiDAR, cameras, LEDs, an IMU, gas and Wi-Fi detectors, and a thermal camera.These sensors are integrated into a shock-absorbing rigid superstructure.
  • III. AU-SPOT MOBILITY SYSTEM: The NeBula Power and Computing Core supplies auxiliary power and computing while protecting internal electronics from falls and supporting extended battery life.It uses an external lithium high-capacity battery and includes custom power-distribution and safety protection.

IV. NEBULA ODOMETRY ON LEGGED SYSTEMS

Reliable odometry is essential for autonomous legged operation in extreme environments, where vibration and degraded perception challenge individual sensing modalities. NeBula addresses this with uncertainty-aware, LiDAR-centric multi-sensor fusion and layered SLAM processing.

  • IV. NEBULA ODOMETRY ON LEGGED SYSTEMS: Reliable odometry is a prerequisite for autonomous operation in environments with darkness, obscurants, self-similar areas, and strong platform vibrations.These conditions make accurate odometry especially difficult on legged platforms.
  • IV. NEBULA ODOMETRY ON LEGGED SYSTEMS: Uneven or slippery surfaces impair inertial sensing and kinematic odometry, while darkness, illumination changes, dust, fog, gas, and visual aliasing challenge cameras.The varied failure modes motivate combining multiple sensing channels.
  • IV. NEBULA ODOMETRY ON LEGGED SYSTEMS: NeBula fuses a selected odometry source as a prior with LiDAR through HeRO, LOCUS, and LAMP to estimate motion and maintain localization.HeRO detects anomalies and selects a reliable odometry input; LOCUS performs LiDAR scan matching; LAMP handles pose-graph and consistency optimization.
  • IV. NEBULA ODOMETRY ON LEGGED SYSTEMS: LOCUS estimates relative motion by minimizing residual error between corresponding points in consecutive LiDAR scans, then refines consistency through scan-to-submap matching.The resulting motion updates the robot pose in the world frame.
  • IV. NEBULA ODOMETRY ON LEGGED SYSTEMS: When HeRO receives no input, the system uses the identity transformation as a prior and reverts to pure LiDAR odometry.This provides a fallback when auxiliary odometry is unavailable.

V. LOCAL PLANNING

Local planning enables Spot to traverse challenging environments by assessing traversability risk and planning safe paths.

  • Local planning assesses traversability risk before generating safe paths for autonomous navigation.

A. Traversability Map

The traversability map represents safe and lethal cells while combining multiple geometric and semantic risk sources across sensing ranges into an aggregated map for action generation.

  • A. Traversability Map: Each grid cell is modeled as safe or lethal, with a probability distribution maintained over the traversability map.The map is a locally 2D grid g=(m1,...,mn), where n=nl×nw.
  • A. Traversability Map: Traversability risks include geometric obstacles and slopes as well as semantic elements such as communication nodes and other robots.
  • A. Traversability Map: Multi-fidelity terrain mapping combines depth cameras, instantaneous LiDAR, and spatially fused point clouds for short-, medium-, and long-range detection.
  • A. Traversability Map: The multi-layer map assigns separate layers to risk sources and applies range-dependent algorithms to detect obstacles and steep slopes.
  • A. Traversability Map: Online evaluation samples sparse cells, then interpolates or inflates their risks to estimate traversability across all grid cells.
  • A. Traversability Map: The aggregated map fuses layer probabilities by element-wise multiplication to produce a conservative risk estimate, illustrated during a real mission.

B. Uncertainty and Perception-aware Planning

Uncertainty- and perception-aware planning predicts how sensing will update traversability beliefs, then selects and repeatedly executes policies that reduce path risk while gathering informative observations.

  • B. Uncertainty and Perception-aware Planning: The planner accounts for noisy, asymmetric, and limited sensors to find short, low-risk trajectories that reduce map uncertainty.
  • B. Uncertainty and Perception-aware Planning: Uncertainty is represented through distributions over cell traversability, with variance capturing confidence in each cell’s mean.
  • B. Uncertainty and Perception-aware Planning: A policy specifies ordered grid locations and robot orientations, while recursive belief updates predict future traversability from sensor characteristics and the planned trajectory.
  • B. Uncertainty and Perception-aware Planning: The optimal policy minimizes total path risk over a K-step path while moving toward a goal.
  • B. Uncertainty and Perception-aware Planning: Cascaded optimization can improve real-time computation by optimizing position and orientation sequentially when predicted-risk computation is challenging.
  • B. Uncertainty and Perception-aware Planning: Receding-horizon execution repeatedly replans from the robot’s new position and favors directions that provide richer sensory input and greater uncertainty reduction.

VI. AREA COVERAGE AND SEARCH BEHAVIOR

The coverage planner represents exploration with a sparse graph and selects actions that trade off coverage information against traversal cost. Coverage belief, entropy, information gain, and expected reward guide policy optimization over frontier and breadcrumb nodes.

  • Mission Objective: The mission objective is to rapidly map and navigate a large unknown environment by maximizing sensor-footprint area covered within the mission time limit.The information-rich graph supports planning under complex topology and terrain.
  • Global Planner: The global planner uses a sparse bidirectional graph whose nodes represent robot poses and whose edges represent connected transitions.Each node stores the probability that its sensor-defined neighborhood has been observed.
  • Graph Construction: Frontier nodes denote incompletely explored regions, whereas breadcrumb nodes denote areas already fully explored.Visiting frontier nodes is expected to change map belief, while breadcrumb visits contribute little new coverage information.
  • Graph Policy: The graph policy maps each node to the next node by maximizing a reward that trades off coverage information and traversal cost.A macro action traverses an edge and updates the graph.
  • Coverage Information: Coverage uncertainty is represented with occupancy Bernoulli distributions over local maps and quantified using entropy.The coverage belief is defined over the graph’s nodes.
  • Coverage Information Gain: Coverage information gain measures the reduction in coverage uncertainty induced by a macro action.The policy objective uses expected future reward, with value-iteration or forward-search methods available for optimization.

VII. EXPERIMENTAL RESULTS

NeBula was field-tested on Spot robots in multi-level, perceptually degraded environments, including DARPA SubT missions. The system combined odometry, traversability, coverage planning, and artifact detection to support autonomous exploration and mapping under operational constraints.

  • Experimental Platform: NeBula was implemented on two Boston Dynamics Spot robots and tested in subsurface, multi-level, GPS-denied environments and industrial power plants.The tested settings included perceptually degraded environments.
  • Urban Circuit Mission: The Urban Circuit deployment combined two Au-Spots, 4 km of exploration, stair traversal, and detection and localization of 16 artifacts, earning team CoSTAR first place.Artifacts included backpacks, survivors, gas leaks, and cellphones detected through multiple sensing modalities.
  • Odometry Estimation: The proposed odometry method produced more accurate maps than KO or KVO across Eagle Rock and NASA JPL environments.KO performed poorly on rough terrain, while KVO was challenged by featureless walls and repetitive carpet textures; their relative performance therefore varied by environment.
  • Traversability: Au-Spot’s perception-aware local planner navigated ramps, slopes, twisted metal, machinery, rubble, narrow corridors, water or oil, stair wells, and raised concrete slabs.The system also successfully avoided negative obstacles such as drop-offs and holes.
  • Coverage Planner: A fleet of two Au-Spots and two wheeled UGVs mapped approximately 25,000 m3 in one one-hour run while traversing four flights of stairs.The resulting map supported artifact localization across multiple levels of an industrial power plant.
  • Complex Mission: During a constrained mission exceeding 1 km, Au-Spot detected and localized artifacts while maintaining localization error below 5 m.The mission included narrow openings, large rooms, raised platforms, rubble, stair descent, and communication-node deployment.

VIII. CONCLUSIONS

The paper presents NeBula on Boston Dynamics’ Spot as a risk- and perception-aware autonomy architecture for complex real-world missions in extreme environments. It reports representative modules and argues that the system advances legged-robot autonomy for live missions.

  • Conclusion: NeBula is a risk- and perception-aware autonomy architecture applied to Boston Dynamics’ Spot for extreme-environment exploration.The paper discusses representative odometry, traversability, and coverage-planning modules for the DARPA Subterranean Challenge.
  • Conclusion: The authors characterize the work as an important step toward legged robots accomplishing complex, real-world, live missions in extreme environments.
Loading 2010.09259v3…