1
00:00:00,250 --> 00:00:04,522
Hand pushes. Foot pushes. Downward pressure.

2
00:00:04,762 --> 00:00:09,130
LUCID repeats a turning motion under added randomized delay.

3
00:00:09,370 --> 00:00:13,474
The full, uncut take runs alongside our story.

4
00:00:15,250 --> 00:00:21,106
Too little randomization leaves policies unprepared. Too much makes learning harder.

5
00:00:21,346 --> 00:00:26,818
LUCID uses learned command–execution feedback to set training difficulty.

6
00:00:28,250 --> 00:00:32,738
Issued commands pass through simulated delay and dynamics.

7
00:00:32,978 --> 00:00:37,490
Command and execution histories enter the same frozen encoder.

8
00:00:37,730 --> 00:00:41,930
Their normalized latent discrepancy guides the next training block.

9
00:00:43,250 --> 00:00:47,714
A denoising V A E learns clean motion from noisy windows.

10
00:00:47,954 --> 00:00:50,570
Then we freeze its temporal encoder.

11
00:00:52,250 --> 00:00:56,834
Bounded P I sets randomization intensity from the latent gap.

12
00:00:57,074 --> 00:00:59,882
Low returns trigger a separate backoff.

13
00:01:00,122 --> 00:01:04,322
One intensity controls six channels for the next training block.

14
00:01:05,250 --> 00:01:09,018
Training uses up to forty milliseconds of added delay.

15
00:01:09,258 --> 00:01:17,178
Frozen policies face sixty milliseconds on full held-out motions, fifty percent above the training maximum.

16
00:01:18,250 --> 00:01:28,522
Under unseen sixty-millisecond delay, completion rises from fifty-two point one percent with filtered-error P I to seventy-three point eight with LUCID.

17
00:01:28,762 --> 00:01:32,194
That is a twenty-one point seven percentage-point gain.

18
00:01:35,250 --> 00:01:41,490
These selected demonstrations combine forty milliseconds of added delay with scheduled pushes.

19
00:01:41,730 --> 00:01:46,146
LUCID is on the left; filtered-error P I is on the right.

20
00:01:46,386 --> 00:01:52,962
They show behavior under perturbation. Aggregate completion comes from separate benchmark trials.

21
00:01:59,250 --> 00:02:03,354
Denoising raises completion over ordinary reconstruction.

22
00:02:03,594 --> 00:02:08,058
Live feedback also improves on one donor’s replayed schedule.

23
00:02:10,250 --> 00:02:14,402
This is the same uncut recording, still at normal speed.

24
00:02:14,642 --> 00:02:22,874
LUCID repeats the turning reference through hand and foot perturbations, with zero-to-sixty-millisecond added randomized delay.

25
00:02:23,114 --> 00:02:32,402
Separately, the paper’s matched forty-millisecond test reports thirty-eight of sixty completions, versus twenty-three for filtered-error P I.

26
00:02:32,642 --> 00:02:36,218
The encoder and curriculum scheduler stay in training.

27
00:02:46,250 --> 00:02:51,002
LUCID uses learned execution feedback to guide domain randomization.

28
00:02:51,242 --> 00:02:55,850
On the robot, only the policy and low-level controller remain.