Skip to content

Tags

Tags

roller-standup-v1-working

Toggle roller-standup-v1-working's commit message
roller-standup: first recipe that actually stands up, both from the b…

…elly AND the back

Training recipe behind wandb run w6rarsdz (4096 envs). Checkpoints 4500 / 5000
are the reference. Anything later that regresses should be diffed against this.

Measured at model_4500, real rolling friction (frictionloss 0.0015, the post-4000
value), 64 envs:

  face-down starts : standing at step 25, tilt 1.0 deg, trunk z 0.1389
  face-up starts   : standing at step 25, tilt 5.1 deg -> 0.7 deg by step 50
  wheel spin during the rise : peak 40.5 rad/s, 90 % of wheels above 1 rad/s
  once standing    : 0.004 rad/s, tilt 0.5 deg, height flat to the millimetre

Run-level, at iteration 5007:

  standing_composite   3.53 / 3.75  = 94 % of max
  ground_state_mix     0.20         (so the score is NOT coming from the
                                     already-standing spawns -- the old failure
                                     signature, composite ~ standing_prob, is gone)
  upright_linear       1.449        -> 13.8 deg of trunk tilt
  height_stand_l1     -0.031        -> 4 mm of height error
  height_progress     +0.030        -> positive, the population gains height

What made it work, in order of discovery:
  1. wheel_support_gate  -- "standing" means standing ON THE WHEELS, killing the
     head tripod that kept 48 % of the task stack without ever getting up
  2. neck_action_rate_l2 dropped -- an attempt tax worth -1.36/step, larger than
     the whole positive task block during discovery
  3. face_up_roll_max = 90 deg -- the walker's built-in reverse curriculum, which
     this env had simply never passed
  4. height_progress (weight 200) -- the potential-based dz that finally gave the
     back start a gradient; without it the whole rollover was worth +0.156/step
     and its first half was downhill

KNOWN LIMITATIONS at this tag, both measured:
  - the rise takes 0.5 s and is violent on the real robot: nothing opposes speed,
    the three anti-violence terms together are 0.18 % of the task reward
  - the standing phase is a perfectly static equilibrium with the wheels frozen,
    a regime the real robot will never get (slope, latency, backlash)