Stanford Robotics Seminar ENGR319 | Autumn 2025 | Make Every Step an Experiment

This research focuses on developing terrain-aware, high-mobility robots for future planetary exploration. Traditional robot-aided exploration faces significant challenges on complex planetary surfaces, such as scattered rocks, steep craters, and loose regolith, which can cause robots to get stuck, halting scientific discovery. The work aims to equip robots with a better understanding of their physical environments and the ability to adapt their operations.

A key innovation is using the robot's own legs as sensors to gather information about terrain properties, such as regolith mechanics, through direct interaction. This "proprioception-based sensing" leverages direct-drive actuators with high force transparency, allowing legs to feel forces from the environment with each step. This contrasts with vision-only approaches, which can be insufficient for discerning critical differences in soil compaction and strength.

The research explores how these force-sensing capabilities can be integrated into legged robots for dynamic locomotion and to map terrain properties with high spatial resolution. This mapping is then used to predict traversal risks for wheeled rovers, enabling safer exploration. Furthermore, the concept of collaborative robots is introduced, where scout robots can identify hazards and potentially assist larger rovers in case of failure, enhancing both safety and scientific discovery.

The Challenge of Planetary Exploration [00:24]

• Robot-aided exploration is essential for planetary missions, but current expectations are higher as humanity prepares for lunar and Martian landings.
• Robots must safely reach destinations, operate reliably across diverse surfaces, and adapt to changing environmental conditions.
• Planetary surfaces are challenging environments with large rocks, steep craters, and unconsolidated loose regolith that can cause robots to slip, sink, or get stuck, leading to costly mission failures and loss of scientific opportunities.
• These challenges extend beyond just locomotion and navigation to operations like excavation and sampling.

To succeed in these planetary services, our robots not only have to reach safely in these destinations but also have to be reliably operate across a diverse range of planetary surfaces and also to adapt their operation flexibly based on the changing environmental conditions.

Sensing Environmental Properties Through Interaction [02:18]

• The speaker's group is developing new sensing, locomotion, and planning tools to enable robots to better understand their physical environments.
• The core questions explored are:
  • How can robots effectively gather information about their environment, especially extracting physical properties through interaction?
  • How can robots be turned into scientific instruments that use their legs to gather information with every step?
  • How can robots predict the outcomes of their interactions based on sensed properties and make informed decisions for exploration, individually or cooperatively?
• The research aims to enable robots to better collaborate and adapt their exploration for improved scientific outcomes.

So in today's talk, I'm going to focus on three key questions that we seek to answer.

Understanding Regolith Mechanics [03:51]

• A primary challenge in unknown environments is gathering information about the surroundings to make informed decisions.
• The research focuses on regolith mechanics and properties, noting that unconsolidated materials on the Moon, Mars, and other planets pose significant challenges for mobility, navigation, drilling, excavation, and construction.
• Assessing regolith properties before navigation or construction is crucial to prevent robots from getting stuck or making poor construction site decisions.
• Visual inspection alone is insufficient to discern critical properties like regolith softness or strength, as visually similar materials can differ greatly in strength due to compaction. A difference of just a few percent in packing fraction can lead to a five-fold difference in strength, which is critical for rover navigation and construction.

However, the challenge is that these properties such as how soft and how much strength the regulus are on the moon or on the Mars can be quite difficult to discern by vision alone.

Proprioception-Based Sensing [05:35]

• Humans and animals use their feet to gather rich, free, and dense information about their environment with every step, which is more informative than vision alone.
• This proprioception-based sensing allows robots to gather information opportunistically while performing other tasks, like transporting material.
• The approach avoids relying on extra sensors to minimize potential failure points.

So for us humans and many of the animals when we are walking our feet is providing information that sometimes can be more informative than vision.

Revolutionizing Robot Legs with Direct Drive Motors [06:37]

• The research utilizes direct-drive motors with high force transparency, enabling robot legs to be highly sensitive to external forces.
• Historically, direct-drive motors had limited power density, leading to the widespread use of gearboxes. However, recent advancements in increased gap gradients have significantly improved torque density, making them suitable for legged mobility.
• This breakthrough allows robot legs to serve not only as locomotive limbs but also as sophisticated sensors and scientific instrumentation for characterizing regolith mechanics in planetary explorations.

So because of this high force transparency, if we make robot legs using these actuators, you can imagine that the leg can be super sensitive to any external forces the leg felt and that can build us robots that can sensibly feel the world from every single interaction.

Traveler: A Single-Legged Sensing Platform [08:43]

• The research began with a single leg from a quadruped robot, mounted on a static stand, utilizing direct-drive actuators for high force transparency.
• This leg, named "Traveler," is a five-bar linkage leg with two coaxial motors.
• The motors can position the toe to perform various sensing protocols, such as poking or scraping the regolith.
• The high force sensing capability allows for the deduction of applied forces by measuring motor current or torque, which can then be translated into information about the regolith's mechanical properties through kinematics.

So John took um he built a single L mounted on static stand using direct drive actuators that has a very high force transparency and he named it after the well-loved USC mascot traveler.

Lab and Field Testing of the Sensing Leg [09:48]

• Two configurations of the leg were created: one mounted on an ergonomic test stand for easy transport to various terrains (snow, desert, hills) for geotechnical testing, and another designed to be mounted on other robots and rovers to decouple locomotion and sensing for data gathering and capability validation.
• A simple sensing protocol involves the toe penetrating the soil and then dragging horizontally.
• The measured forces during penetration reveal the soil's stiffness, while the forces during scraping indicate its shear strength.
• These basic protocols provide rich signals about the material's strength, with potential for developing more active sensing protocols.

And on the bottom plot, I'm showing you the inferred force in the x, which is horizontal direction and also the y in the vertical direction.

Validating Force Sensing Accuracy with a Fluidized Bed [11:47]

• To assess force sensing accuracy, a lab experiment used a fluidized bed.
• A fluidized bed uses airflow to fluidize sand or other planetary simulants, allowing for precise control over the material's strength by adjusting the airflow.
• This setup also allows for easy resetting of the ground surface after disturbances to ensure consistent initial conditions for testing.
• The robotic leg was mounted on the fluidized bed, and its force responses were measured as it poked into the material set at different strengths.
• The results showed a near-linear decrease in sand stiffness with decreasing volume fraction (packing density), consistent with prior measurements and literature, validating the leg's ability to accurately measure surface properties.

So the curve from the top to bottom is when we set the sand strength or the volume fraction which is packing density of sand from the highest to the lowest.

Field Deployments: Wenatchee and Mount Hood [14:21]

• After lab validation, the sensing leg was deployed in two field sites:
  • Wenatchee, New Mexico: A dune field with features like sand ripples and crusty surfaces, similar to those observed on Mars.
  • Mount Hood, Oregon: A site with glaciers and volcanic dust, creating a unique icy regolith mixture relevant to lunar permanently shadowed regions.
• This work is part of a NASA project called LASSIE, involving planetary scientists, roboticists, and researchers studying human-robot interaction and decision-making.

One is Wense in New Mexico which is a dune field that possess very similar features such as the sand ripples and the crusty surfaces that has been observed by the opportunity rover on the Mars.

Distinguishing Icy Regolith and Detecting Bioactivity [15:47]

• At Mount Hood, the leg encountered icy regolith with varying ice content and binding forms.
• The force responses during penetration revealed distinct failure signatures: sudden drops due to surface ice rupture, followed by more ductile failure in cohesive soil with ice bridges.
• These signatures allow for inference of the type of icy regolith and its ice content, which is crucial for understanding lunar exploration and the preservation of water.
• At the Wenatchee analog site, differences in force signatures were observed between active and stabilized dunes.
• The ductile, "leathery" failure in stabilized dunes with vegetation was linked to biopolymers (EPS) emitted by bacterial activity, suggesting that force sensing can reveal signs of past or present bioactivity on other planets.

The implication of that is by looking at these force responses and the resilience of the crust signature, we can potentially reveal signs of either past or present bio activity on different planet.

Integrating Sensing into Dynamic Locomotion [21:35]

• To gather rich data through every step, robots need to walk dynamically while maintaining high-precision force sensing.
• The research integrated this sensing capability into the "Spirit" quadruped robot, which has low gear ratio joints (6:1) offering good force sensing accuracy and sufficient torque density.
• The robot is equipped with differential GPS to record its path trajectory.

So one of the key thing is we need to figure out how to have the robot work in either dynamically or starting slow and eventually dynamically while still maintaining this high precision force sensing measurements.

Developing a Sensing Gait [22:22]

• A sensing gait was developed for the robot to walk while maintaining sensing accuracy, starting with a slow and static gait.
• In this gait, the robot supports its body with three legs and moves one leg at a time, mimicking static field tests. This approach minimizes inertial effects like pitch, roll, and yaw.
• The robot was tested on a track with varying sand compaction and surface crusts, measuring force per unit depth to assess stiffness.
• Initial results showed that the robot could effectively distinguish sandy regions from soft materials, with each leg acting as an individual sensor.

So we started with a relatively slow and static gate. This is a work led by my posto Diego and my amazing another amazing undergrad Ethan.

Real-time Terrain Mapping in Field Deployments [24:01]

• The robot was deployed in planetary analog field sites (Wenatchee and Mount Hood) to measure force during locomotion and create spatial regolith property maps.
• Tests involved traversing from packed to loose sand, where the robot recorded toe position, toe force, and computed sand stiffness.
• Significant decreases in estimated normal strength were observed transitioning from packed to loose sand.
• The force-per-depth curve also revealed surface crust indications.
• At Mount Hood, the robot mapped patchy environments with ice-soil mixtures, generating high spatial resolution maps of regolith strength that captured rapid variations over short distances.

So the top diagram is showing our science team before the rob sending the robot looking at the boundaries visually and determine what are the scientific targets and where should the robot work and the bottom video shows one of the GoPro cameras we mounted on the robot recording every step it takes along with force sensing data.

Vision: Legged Robots as Scouts and Navigational Aids [26:20]

• The vision is to use legged robots as scouts for larger rovers, identifying scientifically valuable sites (e.g., containing water ice or interesting crusts).
• Simultaneously, these scout robots can identify traversal risks for the rovers, helping them avoid soft sand pits or hazardous regions.
• This approach leverages the lower cost of legged robots compared to wheeled rovers to map terrain and ensure safer rover exploration.

Our vision is to one use the robot as a scout in front of rovers to identify um potential scientific valuable sites such as the site that may may contain water ice or contain interesting crusty um signature that of scientific value.

Predicting Rover Traversal Risk [27:17]

• The second challenge is relating mapped regolith strength to the traversal risk for different types of rovers with varying specifications.
• The relationship between regolith strength and traversal risk is not always linear, and there can be critical failure points.
• Understanding the failure mechanisms of robots and rovers in sand is key to predicting these failure points.

So in order to do that one of the first questions we need to ask is why do robots and robers fail on sand? what is the failure mechanism and that will enable us to predict the failure point that lead to those um uh locomotion prop uh locomotion performance degrade.

Sand's Dual Nature: Solid-like vs. Fluid-like [28:24]

• Sand exhibits both solid-like and fluid-like behavior depending on interaction.
• The transition between these behaviors is often critical for successful or failed locomotion.
• A major challenge is the lack of equations to directly relate microscopic particle dynamics to macroscopic responses that impact robot mobility.

You can build a sand castle where you um uh elicit the solid-like material of the sand or you can have the sand flowing through your finger uh which behaves like a fluid.

Locomotion Performance and Yield Stress [29:54]

• Experiments in a fluidized bed showed that geckos and robots exhibited significantly different locomotion behaviors in packed versus soft sand.
• In soft sand, locomotion speed dropped, and sinkage increased, with the creature appearing to "swim" in a fluid-like environment.
• Locomotion success is tightly related to the "yield stress" behavior of the sand.
• Yield stress materials behave rigidly below a certain stress threshold and fluidly above it.

One key thing that we figure out is that the locomotion success is tightly related to the yield stress behavior of the sand.

Predicting Mobility Based on Soil Properties [32:20]

• The yield stress behavior allows for the prediction of robot mobility based on solidification depth alone.
• When sand is rigid, it solidifies at a shallow depth, allowing robots to push against it for propulsion. As sand weakens, larger depths are required for solidification, leading to deeper sinking before effective movement.
• This provides a simple model to predict robot mobility and traversal risk based on the yield strength of the material, sensed by robotic legs.
• Experimental data from various animals and robots collapsed onto a single curve when plotting speed against sensed solidification depth.

This give us a way to very easily determine the performance. For example, if you have you predict a very small significant depth, we the robot can move forward quite effectively.

Validating the Locomotion Model [33:30]

• The locomotion model was validated in lab experiments using the sensing leg to probe different sand sensitivities.
• The model predicted robot speed based on its physical specifications, which was then compared to experimentally measured speeds.
• The model effectively captured the robot's failing trend and identified stiffness levels where the robot began to struggle.
• These validated models were then applied to larger rovers in field analog machines.

With this lab validation we then predict instead of this prototype robot we make this prediction on larger rovers and we validate our model in our two field analog machines.

Collaborative Robot Missions: NASA AMES and Yense [34:43]

• The project aims to enable legged robots and rovers to work in collaborative pairs for safe traversal and recovery from failure.
• Two field missions were conducted:
  • NASA Ames Research Center: Utilized a lunar regolith testbed to validate real-time regolith strength map construction and risk estimation.
  • Wenatchee (Yense) field site: Deployed legged robots and a rover to validate real-time regolith map construction and safe rover traversal based on terrain information.

The goal of this project is to see if we can have leg robot with rover to work in collaboration pairs to enable safe traversal and also to recover from failure.

Lunar Regolith Testbed Experiment [35:45]

• An experiment in a lunar regolith simulant testbed (LHS1) involved a scout legged robot and various rovers with different payload capacities.
• The testbed was prepared with a gradient of regolith strengths: high strength (compacted), intermediate strength (raked), and super soft/fluffy (sifted).
• The scout robot zigzagged across the testbed, sensing terrain strength with each step, and generating a terrain strength map.
• The inferred strength map closely matched the designed pattern, with color indicating strength (red for high, blue for low) and transparency indicating data confidence.

The on the left side the panel A is showing overlay of the designed pattern on the regless test bed and C and D is showing the design pattern and the robot actual measurements.

Traversal Risk Estimation and Path Planning [38:44]

• Based on the terrain map, traversal risk was computed for different rover specifications (payload capacity).
• Higher payload capacities significantly increased the high-risk regions and narrowed the low-risk paths, guiding payload configuration and rover parameter selection.
• At the Wenatchee field site, a naive path (direct route) and a risk-aware path were compared.
• The naive path led to the rover getting stuck in soft sand, while the risk-aware path successfully circumvented hazards and reached the goal.

So we can use this to guide the payload configuration across different rovers. We can also plug in different rover parameter to generate the same traversal risk map.

Scout Robot for Failure Recovery [42:59]

• A secondary aspect of the project explores using scout robots with telescoping arms to assist rovers that get stuck.
• In initial tests, two scout robots docked with a struggling rover and used their arms to push or pull it out of difficulty.
• This demonstrates the potential for scout robots to aid in failure recovery, enhancing mission reliability.

And this is actually the second part of our luster project ma mostly done by Cindy and Mark's group where we wonder if the rover does get stuck if unexpected failure does occur can we use the scout robot to land arm to help them get out of trouble.

Future Vision: Enhanced Science and Failure Prevention [44:14]

• The goal is for scouting robots to provide terrain property estimations and guidance for failure recovery.
• This will enhance scientific discovery and prevent catastrophic failures during planetary exploration.

So with that, we're hoping that going forward, the scouting robot can provide a useful training uh property estimation, but also provide useful guide on failure recovery if the robot gets in trouble to enhance our science and uh prevent the catastrophic failures that we might encounter in during the planetary exploration.

Calibration of Proprioceptive Sensing [44:41]

• Calibration occurs at multiple levels: joint level (torque estimation accuracy) and appendage level (e.g., vertical penetration, hanging weights).
• Manufacturer-specified torque constants are generally accurate.
• Dynamic calibration is more complex due to inertial effects, body orientation, and ground contact detection accuracy, which introduce noise.
• Ongoing research compares different gait speeds, finding that while sand strength can still be reasonably picked up at higher speeds, detecting layering and surface crusts becomes less reliable.
• Over time, joints can degrade due to wear and environmental factors (dust, moisture), leading to performance degradation that requires recalibration of torque and cogging.

So that's actually still in some of our ongoing work where one of the work we are comparing different gate speed.

Optimizing Robot Teams and Hardening Rovers [48:15]

• The optimization involves balancing the use of multiple, less capable robots against fewer, more robust ones.
• The "trust project" explores how teams of robots can collectively perform challenging tasks.
• Having multiple units provides redundancy and allows for more efficient coverage of larger areas.
• The idea is for robots to have individual capabilities, sense their environment, communicate findings, and collaborate for complex tasks, such as one robot anchoring to push another over an obstacle.

So the the kind of one of the idea of this project can we have this kind of team of robot that individually they already have pretty good capability and maybe each of them can carry different sensor payload but the idea is that they are also uh feel the environment every step so they can talk to each other say hey you know I'm in a patch where I think you know you robot who's carrying this uh green size um camera should come over to look at the spot.

Environmental Influences on Sensing [50:33]

• Environmental conditions like wind, humidity, and temperature can influence sensing results.
• At Mount Hood, an icy regolith site, the system is time-varying due to melting and refreezing ice.
• Thermal imaging and temperature changes are correlated with force changes to understand dominant factors influencing surface properties.
• Repeated trials under different environmental conditions (e.g., lunar temperature cycles) can help characterize these influences.

So this is actually something we're currently very excited about is looking at how the force can relate to the environmental condition change and then trying to figure out the dominant factors like you know what's causing the change in surface product because it's such a low cost now you can have the robot take one pass in the in the early in the day you can have it go back and do a repeated trial and see okay what changes right.

Mechanical Design and Environmental Robustness [52:13]

• As robots move towards flight-oriented missions, mechanical design for environmental robustness is critical, balancing sealing with heat dissipation needs.
• Motors can overheat in atmospheres with high temperatures.
• Field trips in harsh environments (e.g., desert heat) highlighted how temperature affects not only human operators but also robots, laptops, and batteries.
• The design challenge involves finding the right balance between sand-proofing, waterproofing, and heat dissipation.

So that's a balance that we need to think very carefully about is how much of the sandproof and waterproof that we need to do as compared to the heat dissipation.

Extending Robot Lifetime Through Planning [53:57]

• Extending robot lifetime involves smart path planning and decision-making, balancing reward and risk within time and battery constraints.
• Current operations rely on human scientists analyzing data and planning rover movements.
• Future robots will need to make complex decisions autonomously based on incoming sensor data and changing environmental conditions.
• Research focuses on converting human decision-making strategies (adapting plans based on gathered information) into algorithms for robots.

A key question is how to balance the reward and risk and then the of course within the time and the battery budget that you have.

Investigating Failure Cases: Slippage and Fine Snow [56:00]

• Failure cases like slippage and performance in fine snow are investigated.
• In icy regolith, stick-slip failures can occur due to ice sheet breakage, leading to rapid entrapment.
• Cohesive media like ice and snow present challenges, as leg extraction can be difficult.
• Different failure modes, such as sinkage failure and slippage failure, depend on factors like penetration angle.
• Researchers strategically interact with the environment (e.g., changing foot design, touchdown angle) to improve mobility.

So essentially, you are stepping in, make ice sheet, break it, fresh snow again, make another ice sheet, break it. It was terrible.

Geometry and Sensing Harmfulness [59:32]

• The geometry of the robot's interaction with the terrain is crucial for sensing effectiveness.
• Research explores using "resistive force theory" to break down arbitrary trajectories into infinitesimal flat plates, each with an orientation and movement direction.
• The robot gathers force responses from these segments, summing them to understand the terrain.
• The design and trajectory of the robot's movement can significantly influence how informative the sensed data is for characterizing terrain properties.
• Future work involves optimizing appendage design, control, and hardware selection based on these trade-offs.

So I can so what we are doing is we are using a representation called resistive force theory. So that essentially breaks down the the arbitry either like shape or le infinite decimal pieces.

Real-time Sensing for Manipulation and Planning [01:01:36]

• The methodology for sensing regolith strength and converting it to traversal risk is being adapted for both wheeled and legged locomotion.
• This involves breaking down contact surfaces (wheels or legs) into infinitesimal flat plates and analyzing their forces to determine propulsion or slippage.
• This approach allows for platform-specific risk estimation based on geometry and motion.

It's the same methodology we plan to apply to both to adapt to general platform specific risk.