Stanford Seminar - From the surface of Mars to the ocean of Enceladus

Robotic Exploration Paradigms

• Robotic exploration has evolved through different paradigms:
  • Exploration 1.0: Fast-paced trial and error missions, successful for the Moon's proximity.
  • Exploration 2.0: Incrementally sophisticated robotic capabilities, successful for Mars.
• Exploration 2.0 may not be suitable for more challenging destinations like icy moons of gas giants.

Robotic Exploration 3.0

• Robotic Exploration 3.0 is proposed with a new adaptive robot called EELS (Exobiology Extended Life Surveyor).
  • EELS is a snake-like robot designed to be highly adaptive, learning from its own experiences and evolving during the mission.
• Exploration 3.0 involves sending a single adaptive robot to an unknown environment, starting with low-complexity tasks and gradually adapting to perform complex tasks.
• The future mission with EELS would involve landing on an icy moon, finding an entrance to a crevasse, exploring the subsurface ocean, collecting samples, and searching for signs of life.

Robot Design and Capabilities

• The robot is viewed as a system of software, hardware, instruments, and ground systems with autonomy, neo-autonomy, and risk-aware autonomy.
• The robot has a modular design with 30 to 48 degrees of freedom, an active skin proposition with rotating screws, and the ability to add payloads anywhere on the body.
• Proceptive control is a key element, allowing the robot to climb without visual feedback by relying on proprioceptive information.
• A version of SLAM from ICP was created, modeling the uncertainty involved in estimation for use by the planner and controller.
• Machine learning was used for acquiring new gate patterns, combining model-free and model-based approaches.

Field Tests and Demonstrations

• Five robots were prototyped, including the 1.0 and 1.5 robots tested in various environments such as ice rinks, the JPL Mars yard, snow mountains, and glaciers.
• The robot demonstrated its versatility by switching to different locomotion modes when encountering obstacles, such as side winding to escape from sand traps.
• The final field test took place at a glacier in Canada, where the robot was tested in various conditions and successfully navigated a moulin, a deep ice channel carved by flowing water.
• The speaker describes a successful test of a robot's vertical mobility in a natural opening.

Risk-Aware Planning and Optimization

• Risk-aware planning involves trading between the level of risk and the desired outcome.
• Joint chance constraint is a mathematical formulation that specifies the probability of violating safety constraints.
• A mathematical trick can decompose the joint chance constraint into manageable individual chance constraints.
• Risk allocation can be viewed as a resource allocation problem where the total risk is distributed over constraints, time, and different places.
• In multi-agent scenarios, risk allocation can be interpreted as a market-based mechanism where agents trade risk at a certain price.
• Dual decomposition converts the primal problem with joint chance constraint into a distributed primal problem and a centralized dual problem.
• The dual problem has only one variable, making it scalable for a large number of agents.

Sequential Decision-Making under Uncertainty

• Mars rover landing and driving is an example of sequential decision-making under uncertainty.
• Co-optimizing each step of the rover's mission can lead to better outcomes compared to optimizing each step separately.
• Chance-constrained nonconvex problems can be solved more efficiently by choosing actions randomly.
• Dual optimization converges exponentially and only takes tens of iterations, making it scalable to large problem sizes.

Challenges and Future Directions

• There is a significant gap between mathematical developments and practical situations in chance-constrained optimal control, especially when distributions are not mathematically favorable or unknown.
• Techniques such as decomposing the distribution into multiple modes and using nonlinear techniques can partially address this gap.

Choice of Snake-Shaped Robot

• The choice of snake-shaped robot for the adaptive strategy was made due to its versatility in adapting to unknown vent widths.

Agile Project Management

• Agile project management and iterative hardware and software changes are crucial for successful real-world robotics projects.