Latest / The Last Observer: Science and Beyond / The Differentiable World Model That Changed Humanoid Robotics Overnight

Transcript

The Differentiable World Model That Changed Humanoid Robotics Overnight · The Last Observer: Science and Beyond

• 0:00
  Speaker 1: Welcome to The Last Observer, a journey through science, history,
• 0:04
  Speaker 1: and the universe's greatest mysteries. From ancient enigmas to the
• 0:08
  Speaker 1: deepest questions at the frontiers of reality, we explore the discoveries, ideas,
• 0:13
  Speaker 1: and mysteries that challenge what we think we know. Stay curious,
• 0:18
  Speaker 1: keep questioning, and let's search for the truth beyond the known.
• 0:27
  Speaker 2: Welcome to The Last Observer Podcast. Today we will talk
• 0:31
  Speaker 2: about a breakthrough that could make training advanced humanoid robots
• 0:34
  Speaker 2: dramatically cheaper and faster than anyone expected just a few
• 0:38
  Speaker 2: years ago.
• 0:39
  Speaker 3: Yeah, I was reading about this from a team at
• 0:41
  Speaker 3: UC Berkeley led by researchers including Qiaoyuan Liao, and the
• 0:46
  Speaker 3: numbers are kind of wild. They call the method SuperMimic,
• 0:50
  Speaker 3: and it was just accepted as an oral presentation at
• 0:52
  Speaker 3: the Conference on Robot Learning this year.
• 0:55
  Speaker 2: SuperMimic. So what is it actually doing in practice for
• 0:58
  Speaker 2: these robots?
• 0:59
  Speaker 3: It is teaching humanoid robots to copy complex human motions
• 1:03
  Speaker 3: like dances, flips, side rolls, broad jumps, and other whole
• 1:07
  Speaker 3: body behaviors, but with way less computer power than the
• 1:10
  Speaker 3: previous best approaches needed.
• 1:12
  Speaker 2: How much less are we talking about here? Give me
• 1:14
  Speaker 2: the specific comparison.
• 1:16
  Speaker 3: They compared it to a previous strong method called Sonic.
• 1:20
  Speaker 3: That one needed 128 GPUs running for 7 full days,
• 1:24
  Speaker 3: which works out to about 21,000 GPU hours. SuperMimic beats
• 1:29
  Speaker 3: its performance with just one GPU in three hours.
• 1:32
  Speaker 2: Wait, 128 versus one? And seven days versus three hours?
• 1:37
  Speaker 2: That is more than a thousand times fewer GPU hours.
• 1:40
  Speaker 2: Is that really accurate?
• 1:42
  Speaker 3: Yes, the paper says over a thousand times reduction in
• 1:45
  Speaker 3: GPU hours. And after a bit longer, about 9.8 hours
• 1:49
  Speaker 3: on two GPUs, it fails on zero of the 129,000
• 1:52
  Speaker 3: plus test motions.
• 1:55
  Speaker 2: That is genuinely surprising to me. I always assumed training
• 1:59
  Speaker 2: these robots to move naturally and fluidly required massive compute
• 2:03
  Speaker 2: clusters that only the biggest labs or companies could afford.
• 2:06
  Speaker 3: That was the bottleneck for a long time in humanoid
• 2:09
  Speaker 3: whole body control. Motion tracking is expensive because you need
• 2:12
  Speaker 3: to practice thousands of different movements over and over again
• 2:15
  Speaker 3: in simulation until the policy gets them right without falling.
• 2:19
  Speaker 2: So how do they crack it? Is it some new
• 2:21
  Speaker 2: magic hardware or a much larger neural network or a
• 2:25
  Speaker 2: completely different approach?
• 2:27
  Speaker 3: No, it is a smarter training method. They built one
• 2:30
  Speaker 3: shared differentiable world model of how the robot dynamics work,
• 2:33
  Speaker 3: and then they train the control policy by imagining short
• 2:36
  Speaker 3: rollouts inside that model and taking analytic gradients through time.
• 2:41
  Speaker 2: A world model meaning a learned simulation of the physics
• 2:44
  Speaker 2: and the robot body response to actions?
• 2:46
  Speaker 3: Yes, exactly. But because it is differentiable... they can back-propagate
• 2:51
  Speaker 3: errors through the imagined future and update the policy much
• 2:54
  Speaker 3: more efficiently than pure trial-and-error reinforcement learning.
• 2:58
  Speaker 2: Okay, that sounds promising on paper. But does it actually
• 3:01
  Speaker 2: work on a real robot, not just in simulation?
• 3:04
  Speaker 3: They transferred the same policy to a physical Unitree G1 humanoid,
• 3:09
  Speaker 3: and it performs diverse and acrobatic motions without any per-motion
• 3:12
  Speaker 3: tuning or extra-fine tuning on the real machine.
• 3:15
  Speaker 2: Acrobatic? Like actual flips and rolls on the real hardware?
• 3:19
  Speaker 3: Yes, flips, side rolls, jumps, and dance-like sequences. The demos
• 3:24
  Speaker 3: show it doing things the previous method struggled with or
• 3:27
  Speaker 3: failed on completely already in simulation.
• 3:29
  Speaker 2: I am curious how stable that transfer is. Sim to
• 3:32
  Speaker 2: reel is usually the hard part for robotics, with all
• 3:35
  Speaker 2: the unmodeled friction, sensor noise, and delays.
• 3:39
  Speaker 3: They say the transfer works, and the videos look solid
• 3:41
  Speaker 3: from what is shown. Of course, it is still research,
• 3:44
  Speaker 3: not a commercial product yet, and more independent testing will
• 3:47
  Speaker 3: be needed to confirm how general it is.
• 3:50
  Speaker 2: Still, if this holds up under more scrutiny, it changes
• 3:53
  Speaker 2: who can experiment with advanced humanoid control in a meaningful way.
• 3:57
  Speaker 3: Right. Smaller teams or universities without huge GPU clusters could
• 4:02
  Speaker 3: start training sophisticated motion policies over a weekend instead of
• 4:06
  Speaker 3: waiting for expensive cluster time to free up.
• 4:08
  Speaker 2: That is the exciting part for me. It feels like
• 4:11
  Speaker 2: a step toward more accessible advanced robotics and more people
• 4:14
  Speaker 2: being able to contribute ideas and test them quickly.
• 4:18
  Speaker 1: Exactly.
• 4:19
  Speaker 3: When the compute barrier drops that sharply, the field can
• 4:22
  Speaker 3: move faster because more experiments become practical for more research
• 4:25
  Speaker 3: groups around the world.
• 4:27
  Speaker 2: I keep thinking about how this fits into the longer
• 4:30
  Speaker 2: history of trying to get robots to move well. Early
• 4:33
  Speaker 2: attempts at reinforcement learning for walking took forever, even for
• 4:36
  Speaker 2: simple gates.
• 4:38
  Speaker 3: Yes, and now we are talking about full-body acrobatics trained
• 4:41
  Speaker 3: in hours. The progress in methods has been just as
• 4:44
  Speaker 3: important as the progress in hardware.
• 4:46
  Speaker 2: It makes me wonder what other bottlenecks will fall next
• 4:49
  Speaker 2: once more people can run these kinds of experiments.
• 4:52
  Speaker 3: The history of AI and robotics has always been about
• 4:55
  Speaker 3: finding better ways to learn from data and experience, and
• 4:58
  Speaker 3: this is a clear example of that pattern continuing.
• 5:01
  Speaker 2: Let me try to understand the method a bit better
• 5:03
  Speaker 2: so I can picture how it actually saves all that compute.
• 5:07
  Speaker 2: They use one differentiable world model for all, the different motions.
• 5:12
  Speaker 3: That is the central insight. Different human motions look very
• 5:16
  Speaker 3: different from the outside, but the underlying robot dynamics, how
• 5:20
  Speaker 3: the body responds to forces, contacts with the ground, and
• 5:24
  Speaker 3: joint torques are largely shared across behaviors.
• 5:27
  Speaker 2: So instead of learning everything from scratch for each new motion,
• 5:31
  Speaker 2: the model captures the common physics once and reuses it
• 5:35
  Speaker 2: for the entire library of movements.
• 5:37
  Speaker 3: Yes. They collect transitions from a regular rigid body simulator,
• 5:42
  Speaker 3: specifically MuJoCo Warp, fit the world model to those log transitions,
• 5:47
  Speaker 3: and keep updating the model while training the policy in
• 5:49
  Speaker 3: a joint loop.
• 5:51
  Speaker 2: And then the policy learns by back-propagating through short imagined
• 5:55
  Speaker 2: rollouts inside that world model.
• 5:58
  Speaker 1: Exactly.
• 5:59
  Speaker 3: Back propagation through time gives much stronger and more informative
• 6:03
  Speaker 3: learning signals than the usual trial-and-error policy gradients that most
• 6:06
  Speaker 3: reinforcement learning methods rely on.
• 6:09
  Speaker 2: I can see why that would be more efficient. You
• 6:12
  Speaker 2: get a direct idea of how small changes in the
• 6:14
  Speaker 2: policy affect the future outcome, rather than just a noisy
• 6:18
  Speaker 2: score at the end of an episode.
• 6:19
  Speaker 3: And the imagined rollouts are also faster than running the
• 6:22
  Speaker 3: full simulator, more than 10 times faster at the batch
• 6:26
  Speaker 3: sizes they use in practice.
• 6:28
  Speaker 2: So you get both better gradients and cheaper rollouts. Those
• 6:32
  Speaker 2: two advantages compound into the huge efficiency gain we are
• 6:35
  Speaker 2: talking about.
• 6:36
  Speaker 3: They do. On a big benchmark of over 129,000 motions,
• 6:41
  Speaker 3: after three hours on one GPU, it already has only
• 6:45
  Speaker 3: 25 failures compared to 179 for the baseline.
• 6:49
  Speaker 2: And those failures include the hard ones like flips and
• 6:51
  Speaker 2: side rolls that the other method could not do at all.
• 6:55
  Speaker 4: Right.
• 6:55
  Speaker 3: The baseline failed every single flip and every side roll.
• 6:59
  Speaker 3: Super Mimic fails none of them after the longer training
• 7:02
  Speaker 3: run on two GPUs.
• 7:04
  Speaker 2: That is impressive coverage. Is the comparison completely fair, though?
• 7:09
  Speaker 2: I heard they use different GPUs and different amounts of
• 7:11
  Speaker 2: motion data.
• 7:12
  Speaker 3: They address that in the paper. The baseline used more
• 7:15
  Speaker 3: motion data, about twice as much, and the GPU performance
• 7:19
  Speaker 3: difference does not come close to explaining a thousand-fold gap
• 7:22
  Speaker 3: in total GPU hours.
• 7:24
  Speaker 2: Okay, so the efficiency gain seems real and not an
• 7:27
  Speaker 2: artifact of the comparison setup.
• 7:30
  Speaker 3: The world model prediction error stays remarkably stable even as
• 7:34
  Speaker 3: you add more motion data, up to hundreds of hours
• 7:36
  Speaker 3: of motion. The compute needed to reach a target performance
• 7:40
  Speaker 3: grows only slowly.
• 7:42
  Speaker 2: That suggests it can scale. You can keep adding more
• 7:45
  Speaker 2: behaviors without the training cost exploding the way it did
• 7:48
  Speaker 2: with pure trial and error methods.
• 7:50
  Speaker 3: Which is what they mean when they say compute is
• 7:53
  Speaker 3: no longer the bottleneck for this kind of whole-body control.
• 7:56
  Speaker 2: I like that framing. It shifts the problem to other
• 7:59
  Speaker 2: things like getting good motion data and testing safely on
• 8:02
  Speaker 2: real hardware.
• 8:03
  Speaker 1: Exactly.
• 8:04
  Speaker 3: The next expensive parts might be collecting high-quality human motion
• 8:08
  Speaker 3: examples and verifying the robots do not fall over or
• 8:11
  Speaker 3: hurt themselves in the real world.
• 8:14
  Speaker 2: It also makes me think about how this kind of
• 8:16
  Speaker 2: model-based approach has been tried before in robotics— but rarely
• 8:20
  Speaker 2: beat the simple policy gradient methods in wall clock time.
• 8:23
  Speaker 3: Yes, model-based reinforcement learning is known for data efficiency, but
• 8:28
  Speaker 3: it has often been slower overall. Their insight about sharing
• 8:32
  Speaker 3: one dynamics model across all motions seems to have flipped that.
• 8:36
  Speaker 2: Seeing it on the real Unitree G1 makes a difference
• 8:39
  Speaker 2: for me. It is not just numbers in a paper
• 8:41
  Speaker 2: or curves in a figure.
• 8:43
  Speaker 3: Yeah, the clips show the same policy doing flips, side rolls,
• 8:47
  Speaker 3: broad jumps, and other acrobatic moves on the physical robot,
• 8:50
  Speaker 3: and they look controlled and repeatable.
• 8:53
  Speaker 2: No special retuning for each motion. That is rare in
• 8:56
  Speaker 2: this area of robotics.
• 8:58
  Speaker 3: Yes, zero-shot transfer in that sense. One general motion tracker
• 9:02
  Speaker 3: that works for all the motions in the library.
• 9:05
  Speaker 2: I wonder how robust it is to different floors or
• 9:07
  Speaker 2: slight wear on the robot joints or unexpected weight shifts.
• 9:11
  Speaker 3: They do not claim it solves every sim-to-real issue. The
• 9:14
  Speaker 3: paper focuses on the training efficiency and the successful transfer demos.
• 9:19
  Speaker 3: The world model even predicts things like sliding on a
• 9:21
  Speaker 3: slippery floor when it is told the friction coefficient.
• 9:24
  Speaker 2: That is a nice touch. It understands the dynamics enough
• 9:28
  Speaker 2: to generalize a bit within the model.
• 9:30
  Speaker 3: And because training is so cheap now, you could iterate
• 9:33
  Speaker 3: quickly if something fails on the real machine. Retrain a
• 9:36
  Speaker 3: new policy in hours instead of waiting days for a
• 9:39
  Speaker 3: big cluster to free up.
• 9:41
  Speaker 2: That feedback loop speed is huge for research progress. Faster
• 9:45
  Speaker 2: cycles mean more ideas get tested and refined.
• 9:48
  Speaker 3: It also opens the door to measuring scaling laws properly.
• 9:51
  Speaker 3: Now that compute is not the main limit, researchers can
• 9:55
  Speaker 3: study how performance grows with more data or better model architectures.
• 9:59
  Speaker 2: Like we saw with large language models, once training got
• 10:02
  Speaker 2: efficient enough, people started mapping the trends and making predictions
• 10:06
  Speaker 2: about future capabilities.
• 10:08
  Speaker 3: Similar idea here for physical intelligence. Humanoid whole body control
• 10:12
  Speaker 3: might follow clearer patterns once more groups can run the experiments.
• 10:16
  Speaker 2: What about other robots? Is this method specific to the
• 10:19
  Speaker 2: Unitree G1, or does it generalize to different body designs?
• 10:24
  Speaker 3: The method is general for humanoid whole-body control. There is
• 10:27
  Speaker 3: already a follow-up work exploring shared models across different robot morphologies,
• 10:32
  Speaker 3: like the Unitree G1 and other designs such as AggieBot
• 10:36
  Speaker 3: or Booster.
• 10:37
  Speaker 2: So potentially one model that works for multiple robot body
• 10:40
  Speaker 2: designs with only small adjustments.
• 10:43
  Speaker 3: Early signs of that. It could reduce the need to
• 10:45
  Speaker 3: retrain from scratch every time the hardware changes or a
• 10:48
  Speaker 3: new robot is released.
• 10:49
  Speaker 2: That would be practical for companies building different sizes of
• 10:52
  Speaker 2: humanoids or for labs that want to test ideas without
• 10:56
  Speaker 2: owning a huge server farm. Yes.
• 10:58
  Speaker 3: And for university groups that previously could not afford the
• 11:01
  Speaker 3: compute to compete in this space?
• 11:03
  Speaker 2: I am also thinking about the data side. Where do
• 11:06
  Speaker 2: the motion examples come from? Motion capture of humans?
• 11:10
  Speaker 3: Yes, large libraries of human motion data are used to
• 11:13
  Speaker 3: train these trackers. The efficiency means you can now explore
• 11:17
  Speaker 3: how to use more of that data effectively.
• 11:19
  Speaker 2: Looking at the bigger picture, this feels like another step
• 11:22
  Speaker 2: in making robots more capable without requiring extreme resources that
• 11:26
  Speaker 2: only a few organizations can.
• 11:28
  Speaker 3: Access. From early reinforcement learning experiments that took days just
• 11:32
  Speaker 3: for simple walking or balancing, to now training thousands of
• 11:36
  Speaker 3: complex whole body motions in hours on one. CARD.
• 11:40
  Speaker 2: The history of robot learning keeps accelerating. What used to
• 11:43
  Speaker 2: be a specialized skill requiring big teams and big budgets
• 11:47
  Speaker 2: is becoming more approachable for more people.
• 11:50
  Speaker 3: Still, there are important caveats. This is primarily motion tracking
• 11:54
  Speaker 3: and control of demonstrated behaviors, not full autonomy or understanding
• 11:59
  Speaker 3: new tasks from natural language instructions.
• 12:02
  Speaker 2: True, it excels at copying and reproducing demonstrated movements smoothly.
• 12:07
  Speaker 2: But inventing new behaviors on the fly or handling completely
• 12:10
  Speaker 2: novel messy real-world surprises is a different and harder challenge.
• 12:14
  Speaker 3: And independent replications will matter a lot. The results look
• 12:17
  Speaker 3: strong in the paper and the demos, but science needs
• 12:20
  Speaker 3: others to try the method and confirm the gains. Agreed.
• 12:24
  Speaker 2: Also, safety testing on physical robots remains critical and cannot
• 12:28
  Speaker 2: be fully replaced by simulation. You cannot just scale that
• 12:32
  Speaker 2: part with more GPUs.
• 12:34
  Speaker 3: Right. The paper focuses on the training efficiency and the
• 12:37
  Speaker 3: transfer demonstrations. Real deployment in homes or factories will need
• 12:41
  Speaker 3: careful validation, risk assessment and probably more work on robustness.
• 12:47
  Speaker 2: Even so, lowering the barrier this much could lead to
• 12:50
  Speaker 2: more experiments, more diverse data and faster progress overall in
• 12:54
  Speaker 2: the field.
• 12:56
  Speaker 3: I can imagine smaller research groups and even startups now
• 12:59
  Speaker 3: contributing meaningful humanoid control work that previously only well-funded labs
• 13:03
  Speaker 3: could attempt.
• 13:04
  Speaker 2: That diversity of ideas and approaches might be the real
• 13:07
  Speaker 2: long-term impact of a breakthrough like this.
• 13:10
  Speaker 3: Exactly. When compute stops being the main gatekeeper, creativity, careful testing,
• 13:16
  Speaker 3: and good data become the main limits, which is a
• 13:18
  Speaker 3: healthier place for the field to be.
• 13:21
  Speaker 2: It is a good reminder that efficiency breakthroughs and algorithms
• 13:24
  Speaker 2: can open whole areas of research that were previously closed off.
• 13:28
  Speaker 3: And make the future of robotics a bit more distributed,
• 13:31
  Speaker 3: a bit more interesting, and potentially a bit faster.
• 13:35
  Speaker 2: Well, that was a fascinating development to dig into together.
• 13:38
  Speaker 2: SuperMimic shows that smarter methods can sometimes beat brute force
• 13:41
  Speaker 2: compute by a huge margin.
• 13:43
  Speaker 3: Yes. One GPU, three hours, and a humanoid doing controlled
• 13:47
  Speaker 3: flips and rolls. The field of robot learning is moving
• 13:50
  Speaker 3: fast right now.
• 13:52
  Speaker 2: Thanks for exploring this with me. It gives a lot
• 13:54
  Speaker 2: to think about for the next wave of humanoid capabilities
• 13:58
  Speaker 2: and who gets to shape them.
• 13:59
  Speaker 1: Definitely.
• 14:00
  Speaker 3: The combination of better algorithms and more accessible training could
• 14:04
  Speaker 3: accelerate progress more than either alone.
• 14:07
  Speaker 2: I am looking forward to seeing how others build on
• 14:10
  Speaker 2: this idea in the coming months.
• 14:12
  Speaker 3: Me too. See you next time.
• 14:14
  Speaker 2: One thing that stands out is how this method removes
• 14:17
  Speaker 2: compute as the main barrier specifically for whole-body motion tracking.
• 14:21
  Speaker 3: Yeah, and that allows researchers to focus on other important
• 14:24
  Speaker 3: questions like how to generate better motion data or how
• 14:28
  Speaker 3: to make the policies more robust to real-world variations.
• 14:32
  Speaker 2: I also wonder about the broader implications for other areas
• 14:35
  Speaker 2: of robotics. Could similar differentiable world model ideas speed up
• 14:39
  Speaker 2: training for manipulation or multi-robot systems?
• 14:43
  Speaker 3: It is possible. The idea of amortizing learning across related
• 14:47
  Speaker 3: tasks with a shared model has been powerful in other domains,
• 14:51
  Speaker 3: and this shows it can work for high-dimensional continuous control.
• 14:55
  Speaker 2: Looking back at the history of AI and robotics …
• 14:57
  Speaker 2: we have seen several waves where better algorithms suddenly made
• 15:01
  Speaker 2: previously impractical experiments routine.
• 15:04
  Speaker 3: From the early days of Q-learning and policy gradients that
• 15:07
  Speaker 3: were slow and sample inefficient to modern deep reinforcement learning
• 15:11
  Speaker 3: and now these model-based approaches that combine the best of both.
• 15:15
  Speaker 2: It is interesting to think about what the next bottleneck
• 15:18
  Speaker 2: will be once training efficiency is no longer the limiting factor.
• 15:22
  Speaker 3: Probably high-quality data, reliable sim-to-real transfer for more complex environments,
• 15:27
  Speaker 3: and safe ways to test and deploy these systems in
• 15:30
  Speaker 3: the real world.
• 15:31
  Speaker 2: And perhaps the ability to combine this kind of low-level
• 15:34
  Speaker 2: control with higher-level planning and language understanding.
• 15:38
  Speaker 3: Yes, so that a robot can not only execute a
• 15:41
  Speaker 3: flip when shown one, but decide when and why to
• 15:43
  Speaker 3: perform different movements based on a goal.
• 15:46
  Speaker 2: That integration is where a lot of the remaining hard
• 15:49
  Speaker 2: problems live.
• 15:50
  Speaker 3: Definitely. SuperMimic solves one important piece of the puzzle very efficiently,
• 15:55
  Speaker 3: which is valuable on its own.
• 15:57
  Speaker 2: I keep coming back to the accessibility angle. When a
• 16:00
  Speaker 2: method like this becomes available, more research groups can start
• 16:03
  Speaker 2: contributing their own motion libraries or testing on their own robots.
• 16:08
  Speaker 3: Yes, and that diversity could lead to better coverage of
• 16:11
  Speaker 3: different body types, different environments, and different cultural or practical
• 16:16
  Speaker 3: movement styles.
• 16:18
  Speaker 2: It also lowers the cost of failure. If an experiment
• 16:21
  Speaker 2: does not work, you have not wasted a huge amount
• 16:24
  Speaker 2: of cluster time.
• 16:25
  Speaker 3: That is important for scientific progress. Faster iteration and lower
• 16:30
  Speaker 3: cost of trying new ideas usually leads to more discoveries.
• 16:34
  Speaker 2: Of course, we should stay measured. The paper demonstrates strong
• 16:38
  Speaker 2: results on motion tracking and some real robot demos, but
• 16:41
  Speaker 2: full general-purpose humanoid capability is still a long way off.
• 16:45
  Speaker 3: Absolutely. This is a significant efficiency improvement for one crucial component,
• 16:50
  Speaker 3: not a complete solution to all of robotics.
• 16:54
  Speaker 2: Still, progress on the hard parts compounds. Better control makes
• 16:58
  Speaker 2: higher-level intelligence more useful, and vice versa.
• 17:01
  Speaker 3: And open sharing of methods like this, with code and
• 17:04
  Speaker 3: project pages, helps the whole community move forward together.
• 17:09
  Speaker 2: I am curious to see the independent replications and the
• 17:12
  Speaker 2: follow-up work that is already appearing.
• 17:14
  Speaker 3: Me too. The fact that a follow-up exploring multi-morphology models
• 17:18
  Speaker 3: appeared so quickly shows how rapidly ideas can spread when
• 17:22
  Speaker 3: the barriers are low.
• 17:24
  Speaker 2: Well, this has been a really interesting development to think
• 17:26
  Speaker 2: through together.
• 17:28
  Speaker 1: It has.
• 17:29
  Speaker 3: SuperMimic is a clear example of how algorithmic ingenuity can
• 17:33
  Speaker 3: unlock capabilities that previously required enormous resources.
• 17:37
  Speaker 2: Thanks for walking through the details and the implications with me.
• 17:41
  Speaker 3: Anytime. Until next time.
• 17:43
  Speaker 2: Before we wrap up, I want to highlight how this
• 17:45
  Speaker 2: work fits into the ongoing effort to make physical AI
• 17:48
  Speaker 2: more practical.
• 17:49
  Speaker 3: Yeah, humanoid robots have been a long-term goal in the field.
• 17:53
  Speaker 3: and efficient training methods like SuperMimic help close the gap
• 17:56
  Speaker 3: between simulation, progress, and real-world usefulness.
• 18:01
  Speaker 2: The fact that the same policy transfers to a real
• 18:03
  Speaker 2: robot without promotion tuning is particularly encouraging for future applications.
• 18:08
  Speaker 3: It suggests that the learned dynamics capture enough of the
• 18:11
  Speaker 3: real physics to be useful beyond pure simulation.
• 18:15
  Speaker 2: The open nature of the research, with project pages and demos,
• 18:18
  Speaker 2: helps accelerate the whole community.
• 18:21
  Speaker 4: Overall, this is a strong example of how targeted algorithmic
• 18:30
  Speaker 4: improvements can have outsized effects on what is possible.
• 18:34
  Speaker 3: Agreed. It is worth following how this and similar ideas
• 18:37
  Speaker 3: develop in the coming months and years.
• 18:40
  Speaker 2: Thanks again for the conversation. This was a great topic
• 18:43
  Speaker 2: to explore.
• 18:44
  Speaker 3: It was. Take care.
• 18:46
  Speaker 2: To put the numbers in perspective one more time, the
• 18:49
  Speaker 2: previous method required roughly 21,000 GPU hours, while SuperMimic reaches
• 18:55
  Speaker 2: better coverage in just a few hours on one or
• 18:57
  Speaker 2: two GPUs.
• 18:59
  Speaker 3: That scale of improvement is rare and worth paying attention
• 19:02
  Speaker 3: to in any field.
• 19:04
  Speaker 2: It also shows the value of rethinking the training process
• 19:07
  Speaker 2: rather than simply scaling up existing methods.
• 19:10
  Speaker 3: Yes, sometimes the biggest gains come from a better understanding
• 19:13
  Speaker 3: of the problem structure, in this case the shared dynamics
• 19:16
  Speaker 3: across motions.
• 19:17
  Speaker 2: For anyone following humanoid robotics or robot learning, this is
• 19:21
  Speaker 2: a development that could influence how future systems are trained.
• 19:25
  Speaker 3: And it is a good reminder that the field is
• 19:27
  Speaker 3: still full of opportunities for clever ideas to have large impacts.
• 19:31
  Speaker 2: We will be watching to see how this approach is
• 19:34
  Speaker 2: used and extended.
• 19:35
  Speaker 1: Same here.
• 19:36
  Speaker 3: Thanks for listening.
• 19:37
  Speaker 2: And with that, we come to the end of this
• 19:39
  Speaker 2: episode of the Last Observer podcast.
• 19:42
  Speaker 3: Thanks for joining us as we explored how SuperMimic is
• 19:45
  Speaker 3: changing the economics of training humanoid robots.
• 19:48
  Speaker 2: We hope you found the conversation as engaging as we did.
• 19:52
  Speaker 2: Until next time.
• 19:53
  Speaker 3: See you then.
• 19:54
  Speaker 2: One final thought. The rapid appearance of follow-up work building
• 19:58
  Speaker 2: on SuperMimic shows how quickly the research community can build
• 20:01
  Speaker 2: on efficient methods.
• 20:03
  Speaker 3: Yes, when the training cost drops, the rate of progress
• 20:06
  Speaker 3: in related ideas can increase as well.
• 20:09
  Speaker 2: That is the kind of compounding effect that makes these
• 20:12
  Speaker 2: efficiency breakthroughs so important.
• 20:14
  Speaker 3: Indeed. It is an exciting time for anyone interested in
• 20:18
  Speaker 3: the future of humanoid robots and physical AI.
• 20:22
  Speaker 2: We will leave it there for today. Thank you for
• 20:24
  Speaker 2: listening to The Last Observer podcast.
• 20:27
  Speaker 3: Thank you.