Escape Without a Recipe: Maneuver-Agnostic Mars Rover Recovery from Granular Entrapment

Key points

• We introduce a recurrent policy that learns to escape without a predefined maneuver.1
• The framework rests on a verified trap: in a continuum material point method (MPM) simulator we identify a configuration at sinkage-to-radius ratio z/r = 2.0 where the strongest of ten scripted controllers succeeds in only 7/32 trials, yet the regolith retains traction and continues to shear, confirming granular entrapment rather than a numerical artifact.1
• A separate simulator mobility latch gates training rewards.1
• Across five training seeds, mean action evaluation achieves 62.5% escape (range 28.1-81.3%; 32 trials per seed) versus 7/32 for the strongest scripted controller and 0/32 for action noise.1

Sources (1)

• [1]Escape Without a Recipe: Maneuver-Agnostic Mars Rover Recovery from Granular Entrapment
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 07:20 PM

  Wheeled planetary rovers immobilised in loose regolith have no principled recovery controller: scripted maneuvers are terrain-specific, and naive driving deepens entrapment.

  We introduce a recurrent policy that learns to escape without a predefined maneuver.

Extractive summary: sentences quoted from the sources.