Skip to main content

Terrain Adaptation & Push Recovery

Duration: 50 min · Level: Advanced · Module: 3. Bipedal Locomotion & Whole-Body Control · Focus: terrain, push-recovery, footstep-planning, robustness

Learning objectives

By the end of this lesson you will be able to explain and apply:

  • Terrain estimation
  • Footstep planning
  • Push recovery strategies
  • MIT Terrain-Adaptive Atlas (2023)
  • Fall detection and protective response

Why this matters​

Real-world environments are unforgiving: wet floors, loose rugs, uneven tile, staircases, and unexpected perturbations from humans.

Overview​

Real-world environments are unforgiving: wet floors, loose rugs, uneven tile, staircases, and unexpected perturbations from humans. A robot that works in a hospital must handle all of these without falling. This lesson covers terrain perception, adaptive footstep planning, and recovery controllers.

Key concepts​

Key idea

Terrain estimation: fuse depth camera + IMU + proprioceptive foot contact to build real-time local elevation map at 100Hz update rate

  • Footstep planning: project safe landing zones from elevation map; use A* or MPC to plan footstep sequence that avoids obstacles and respects terrain slope limits
  • Push recovery strategies: ankle strategy (small perturbation → ankle torque), hip strategy (larger → hip flexion), step strategy (large → take a step to new support polygon)
  • MIT Terrain-Adaptive Atlas (2023): trained RL policy on diverse terrain in simulation; zero-shot deployment across grass, gravel, stairs, and slippery surfaces
  • Fall detection and protective response: detect impending fall at >500ms before impact; robot enters protective posture — protect head, distribute impact across large body areas
  • G1 specification: must withstand lateral push of 150N at hip height without falling; must recover from unexpected 20° slope at 1 m/s walking speed

Check your understanding​

Try to recall each answer before expanding it.

Q1. What do you know about Terrain estimation?

fuse depth camera + IMU + proprioceptive foot contact to build real-time local elevation map at 100Hz update rate

Q2. What do you know about Footstep planning?

project safe landing zones from elevation map; use A* or MPC to plan footstep sequence that avoids obstacles and respects terrain slope limits

Q3. What do you know about Push recovery strategies?

ankle strategy (small perturbation → ankle torque), hip strategy (larger → hip flexion), step strategy (large → take a step to new support polygon)

Q4. What do you know about MIT Terrain-Adaptive Atlas (2023)?

trained RL policy on diverse terrain in simulation; zero-shot deployment across grass, gravel, stairs, and slippery surfaces

Q5. What do you know about Fall detection and protective response?

detect impending fall at >500ms before impact; robot enters protective posture — protect head, distribute impact across large body areas

← Previous: 3.4 Whole-Body Control: Moving & Working Simultaneously

Part of Module 3: Bipedal Locomotion & Whole-Body Control.