Mathematics of Humanoid Robotics
Challenges and Future Research

From Variational Inference, Diffusion Models, Singular Perturbations to Reinforcement Learning

Abhishek Gupta

Associate Professor, ECE | The Ohio State University
Co-Director, IITB-OSU Frontier Center
Founder, Ensemble Control Inc. (https://ensemblecontrol.com/)

The Humanoid Frontier (2023–2026)

Humanoid robotics moved from research demos to funded industrial programs:

Lab Flagship stack Action decoder Motor learning
Physical Intelligence \(\pi_0\), \(\pi_{0.5}\), \(\pi^*_{0.6}\) + Recap flow matching BC \(\to\) RL
Figure AI Helix regression head BC (500 h teleop)
Skild AI Skild Brain omni-body policy large-scale RL + video
NVIDIA GR00T N1 diffusion transformer BC: real + synthetic
BD \(\times\) TRI Atlas LBM (450M) diffusion/flow DiT BC (whole-body teleop)
Google DeepMind Gemini Robotics tokenized decoder BC on VLM backbone
1X Redwood + 1XWM VLA (~160M params) BC + world-model eval

Thesis of this talk

Strip the branding: the whole field runs on six classical mathematical ideas.

Different Vocabularies, Same Mathematics

The labs publish under different names — “vision-language-action models,” “large behavior models,” “robot brains,” “world models” — but the machinery reduces to:

  1. Variational inference & generative modeling — every action decoder (behavior cloning, diffusion policies, flow matching)
  2. On-line parameter estimation — cross-embodiment adaptation (Skild’s rapid motor adaptation)
  3. Mirror descent — the RL layer (PPO, advantage-weighted updates, Recap)
  4. Stochastic robust optimization — sim-to-real via domain randomization
  5. Singular perturbation — dual-system (slow/fast) architectures
  6. KL projection — every teacher-student distillation step

This talk is organized around the mathematics, not the companies.

The Common Architecture

Every frontier stack is the same triple:

\[ \underbrace{h_t=\Phi(o_{t},\ell)}_{\substack{\text{semantic encoder}\\ \text{(vision-language model)}}} \qquad \underbrace{\hat z_t=\phi(x_{t-k:t},\,a_{t-k:t})}_{\substack{\text{adaptation /}\\ \text{estimation module}}} \qquad \underbrace{a_{t:t+H}\sim p_\psi(\cdot\mid h_t,\hat z_t)}_{\substack{\text{generative}\\ \text{action head}}} \]

  • A pretrained VLM backbone \(\Phi\) turns pixels \(o_t\) + language \(\ell\) into task semantics \(h_t\)
  • An estimator \(\phi\) infers latent dynamics from recent state-action history
  • A conditional generative model \(p_\psi\) emits an action chunk \(a_{t:t+H}\), typically \(H \approx 16\)\(50\) steps

The three questions that differentiate the labs

(a) Which generative family implements \(p_\psi\)?   (b) Imitation or reinforcement?   (c) How are the slow and fast modules coupled?

Notation

Symbol Meaning
\(o_t,\ \ell,\ x_t,\ a_t\) camera observation, language instruction, robot state, motor action
\(a_{t:t+H}\) (or just \(a\)) action chunk: the next \(H\) actions, generated together
\(h_t = \Phi(o_t,\ell)\) semantic latent from the vision-language model
\(p_\psi(a \mid h)\) the learned action distribution (“policy head”) with parameters \(\psi\)
\(\pi_\theta(a \mid x)\) RL policy with parameters \(\theta\)
\(\xi,\ z = e(\xi),\ \hat z_t\) true dynamics parameters, their encoding, the on-line estimate
\(\epsilon \sim \mathcal N(0, I)\) standard Gaussian noise
\(\mathrm{KL}(q\,\Vert\,p) = \mathbb{E}_q\!\left[\log \tfrac{q}{p}\right]\) Kullback–Leibler divergence (always \(\ge 0\); \(=0\) iff \(q = p\))
\(A^\pi(x,a) = Q^\pi(x,a) - V^\pi(x)\) advantage: how much better action \(a\) is than average

One convention throughout: \(\psi\) = generative-head parameters, \(\theta\) = RL-policy parameters, \(\phi\) = estimator.

Part I — Action Generation as Variational Inference

Behavior Cloning is Maximum Likelihood

Every frontier system is pretrained on teleoperation data \(\mathcal D = \{(o_i, \ell_i, a_i)\}\) by behavior cloning (BC):

\[ \min_\psi\ \mathbb{E}_{\mathcal D}\bigl[-\log p_\psi(a \mid o, \ell)\bigr] \;=\; \min_\psi\ \mathrm{KL}\bigl(p_{\mathrm{data}}\ \Vert\ p_\psi\bigr) + \text{const} \]

Why the identity holds: \[ \mathrm{KL}(p_{\mathrm{data}} \Vert p_\psi) = \underbrace{\mathbb{E}_{p_{\mathrm{data}}}[\log p_{\mathrm{data}}]}_{\text{entropy: no } \psi} \;-\; \mathbb{E}_{p_{\mathrm{data}}}[\log p_\psi] \]

so minimizing negative log-likelihood \(\equiv\) projecting the demonstrator onto the model family (an M-projection).

Who uses this

Figure trains Helix on \(\sim\!500\) h of teleop exactly this way; Tesla’s Optimus and 1X’s Redwood are the same recipe at different scales. The architectural race is over which family makes this expressive enough.

Why Simple Regression Heads Fail: Mode Averaging

Demonstrations are multimodal: to pass an obstacle, half the demos go left, half go right.

Unimodal Gaussian head \(\;p_\psi = \mathcal N(\mu_\psi(x), \sigma^2 I)\)

MLE fits the conditional mean: \[ \mu_\psi(x) \to \mathbb{E}[a \mid x] = \tfrac{1}{2}a_{\text{left}} + \tfrac{1}{2}a_{\text{right}} \]

\(\Rightarrow\) the robot drives into the obstacle. The average of two valid actions need not be valid.

Generative head (diffusion / flow / autoregressive)

Represents the full conditional distribution: \[ p_\psi(a \mid x) \approx \tfrac{1}{2}\delta_{a_{\text{left}}} + \tfrac{1}{2}\delta_{a_{\text{right}}} \]

\(\Rightarrow\) samples commit to one coherent mode per rollout.

This single failure mode explains the entire move to generative action heads — and generative heads are trained by variational inference.

Variational Inference in Three Lines

Setup. A latent-variable model \(p(a, z) = p(a \mid z)\, p(z)\). The likelihood \(p(a) = \int p(a \mid z)\, p(z)\, \mathrm{d}z\) is intractable.

Idea. Introduce any distribution \(q(z)\) and use Jensen’s inequality (\(\log\) is concave):

\[ \log p(a) = \log \mathbb{E}_{q(z)}\!\left[\frac{p(a, z)}{q(z)}\right] \;\ge\; \underbrace{\mathbb{E}_{q(z)}\!\left[\log \frac{p(a, z)}{q(z)}\right]}_{\textbf{ELBO}(q)} \]

How tight? Exactly the gap to the true posterior:

\[ \log p(a) \;=\; \mathrm{ELBO}(q) \;+\; \mathrm{KL}\bigl(q(z)\ \Vert\ p(z \mid a)\bigr) \]

The variational principle

Maximizing the ELBO (i) pushes up the data likelihood and (ii) pulls \(q\) toward the true posterior — every generative action head in this talk instantiates this bound.

The ELBO, Rearranged for Practice

Split the joint \(p(a,z) = p(a\mid z)\,p(z)\) inside the bound:

\[ \mathrm{ELBO}(q) = \underbrace{\mathbb{E}_{q(z)}\bigl[\log p(a \mid z)\bigr]}_{\text{reconstruction}} \;-\; \underbrace{\mathrm{KL}\bigl(q(z)\ \Vert\ p(z)\bigr)}_{\text{regularization toward the prior}} \]

  • Amortized VI: learn one network \(q_\phi(z \mid a)\) that maps each datapoint to its approximate posterior — one forward pass replaces per-datapoint optimization
  • This is the VAE objective; with a sequence of latents it becomes a world model (later slide)
  • With the latents chosen as progressively noised copies of the data, it becomes a diffusion model (next slide)

Diffusion is not a separate theory — it is the ELBO applied to a particularly clever latent hierarchy.

Diffusion Models I: The Forward (Noising) Process

Model the action chunk \(a^0 \equiv a_{t:t+H}\). Define latents \(a^1, \dots, a^S\) by gradually destroying \(a^0\) with Gaussian noise:

\[ q(a^s \mid a^{s-1}) = \mathcal N\!\bigl(\sqrt{1-\beta_s}\; a^{s-1},\ \beta_s I\bigr), \qquad s = 1, \dots, S \]

Gaussians compose, so the marginal has closed form. With \(\alpha_s = 1-\beta_s\), \(\bar\alpha_s = \prod_{r\le s}\alpha_r\):

\[ a^s = \sqrt{\bar\alpha_s}\, a^0 + \sqrt{1-\bar\alpha_s}\,\epsilon, \qquad \epsilon \sim \mathcal N(0, I) \]

  • \(\bar\alpha_S \approx 0\): after \(S\) steps the chunk is indistinguishable from pure noise
  • The forward chain is fixed — no learning. It plays the role of the variational posterior \(q\)
  • Generation = learning to run this chain in reverse: noise \(\to\) valid action chunk

Key design trick: because \(q\) is fixed by construction, the only thing left to learn is the reverse denoiser.

Diffusion Models II: ELBO \(\Rightarrow\) Denoising Score Matching

The noising chain defines a hierarchical latent-variable model; write its ELBO and it telescopes into per-step KL terms between Gaussians. After the standard reparameterization, the bound collapses to a regression:

\[ \min_\psi\ \mathbb{E}_{s,\,a^0,\,\epsilon}\; \Bigl\|\epsilon - \epsilon_\psi\bigl(\underbrace{\sqrt{\bar\alpha_s}\,a^0 + \sqrt{1-\bar\alpha_s}\,\epsilon}_{\text{noised chunk } a^s},\; s,\; h_t\bigr)\Bigr\|^2 \]

“Given a noised action chunk, predict the noise that was added” — conditioned on the semantic latent \(h_t\).

Connection: score matching

The optimal denoiser estimates the score: \(\ \epsilon_\psi \propto -\nabla_a \log q_s(a^s \mid h_t)\). Training the ELBO \(=\) learning the gradient field of the action distribution; sampling integrates the reverse SDE/ODE along it. And \(\nabla\log q\) represents sets of valid grasps/gaits where a regression head averages them.

Diffusion Policies on Real Humanoids

Who ships it

  • TRI’s diffusion policy \(\to\) action head of the Boston Dynamics Atlas large behavior model (450M-param diffusion transformer)
  • NVIDIA GR00T N1: diffusion transformer as the fast “System 1” at 120 Hz
  • Both condition on VLM latents and emit whole-body action chunks

Why it wins / what it costs

  • ✔ Multimodal: represents many valid behaviors at once
  • ✔ Scales predictably with compute (DiT architecture)
  • ✘ Sampling = integrating the reverse chain: many network evaluations per chunk
  • ✘ Latency is the enemy of high-rate control

The latency problem motivates the next idea

If diffusion’s reverse chain is too slow for 50 Hz control, straighten the paths: flow matching.

Flow Matching: Transport Instead of Denoising

Physical Intelligence’s \(\pi_0\): a 3B PaliGemma VLM + a 300M flow-matching “action expert.”

Draw \(\tau \in [0,1]\), noise \(\epsilon \sim \mathcal N(0,I)\), and connect noise to data by a straight line:

\[ a^\tau = \tau\, a + (1-\tau)\,\epsilon \qquad\Longrightarrow\qquad \frac{\mathrm{d} a^\tau}{\mathrm{d}\tau} = a - \epsilon \]

Train a network to regress this velocity field, conditioned on semantics \(h_t\):

\[ \min_\psi\ \mathbb{E}_{\tau,\,a,\,\epsilon}\; \bigl\| v_\psi(a^\tau, \tau, h_t) - (a - \epsilon) \bigr\|^2 \]

Inference: integrate the ODE \(\dot a^\tau = v_\psi(a^\tau, \tau, h_t)\) from noise \(\epsilon\) at \(\tau{=}0\) to an action chunk at \(\tau{=}1\) — in a handful of steps (\(\sim\!10\times\) fewer evaluations than diffusion).

That factor of 10 is exactly what makes 50 Hz chunked control feasible on-robot.

Diffusion vs. Flow Matching: One Geometry, Two Paths

Diffusion Flow matching
Latent path noise \(\to\) data curved (variance-preserving SDE) straight line
Training target added noise \(\epsilon\) (score) velocity \(a - \epsilon\)
Theoretical frame ELBO / denoising score matching continuous normalizing flow / optimal transport
Sampling many-step reverse SDE/ODE few-step ODE
Frontier users Atlas LBM, GR00T N1 \(\pi_0\), \(\pi_{0.5}\), Atlas LBM (FM objective)

Connection: optimal transport

The straight-line interpolant makes the target field the displacement interpolation of optimal transport between the noise and data measures. Flow matching is the OT-geodesic counterpart of the diffusion ELBO — same endpoints, shortest path.

Autoregressive Tokenization: Discretize and Do MLE

Google DeepMind’s RT-2 / Gemini Robotics line and \(\pi_0\)-FAST take a third route: make actions look like language.

  1. Compress the chunk: FAST applies a discrete cosine transform to \(a_{t:t+H}\), then quantizes — decorrelating and shortening the token sequence (a rate–distortion step)
  2. Train the same next-token cross-entropy as the language model:

\[ \min_\psi\ \mathbb{E}_{\mathcal D}\Bigl[-\textstyle\sum_j \log p_\psi\bigl(u_j \mid u_{<j},\, h_t\bigr)\Bigr], \qquad u = \mathrm{tokens}(a_{t:t+H}) \]

— exactly behavior cloning MLE again, with an autoregressive categorical generative family.

The trade

Gain: actions share the VLM’s vocabulary \(\Rightarrow\) co-training with internet-scale data. Cost: expressivity is bounded by the tokenizer, and generation loops token-by-token at chunk rate.

World Models: The ELBO Made Explicit

1X trains a video world model (1XWM) as a learned simulator; NVIDIA’s GR00T generates synthetic “dream” trajectories. The underlying object is the latent state-space model ELBO over trajectories:

\[ \log p(o_{1:T} \mid a_{1:T}) \;\ge\; \sum_{t=1}^{T}\; \underbrace{\mathbb{E}_{q}\bigl[\log p(o_t \mid z_t)\bigr]}_{\text{reconstruct observations}} \;-\; \underbrace{\mathrm{KL}\bigl(q(z_t \mid z_{t-1}, a_t, o_t)\ \Vert\ p(z_t \mid z_{t-1}, a_{t-1})\bigr)}_{\text{latent dynamics must predict the posterior}} \]

  • \(p(z_t \mid z_{t-1}, a_{t-1})\): the learned dynamics prior — this is the simulator
  • \(q(z_t \mid \cdot)\): the amortized posterior (filtering distribution)
  • Same reconstruction \(-\) KL structure as the VAE slide, unrolled through time

Use case: score policy checkpoints inside the world model instead of physical A/B tests — an estimated off-policy evaluator.

Part II — Embodiment Adaptation as Parameter Estimation

Rapid Motor Adaptation (the Skild Lineage)

The cleanest estimation-theoretic idea in the field (RMA \(\to\) extreme parkour \(\to\) Skild Brain).

Phase 1 (simulation, privileged). Train a policy conditioned on the true environment/embodiment parameters \(\xi\) — mass, friction, terrain, motor strength — through a low-dimensional encoding \(z = e(\xi)\): \[ \pi\bigl(a \mid x_t,\ z\bigr), \qquad z = e(\xi) \]

Phase 2 (deployable). The robot can’t see \(\xi\). Train an adaptation module to recover \(z\) from what it can measure — its own proprioceptive history:

\[ \hat z_t = \phi\bigl(x_{t-k:t},\ a_{t-k:t}\bigr), \qquad \min_\phi\ \mathbb{E}_{\xi}\,\mathbb{E}_{\text{rollout}}\ \bigl\|\hat z_t - e(\xi)\bigr\|^2 \]

What this is, classically

Textbook on-line system identification, amortized: the regression is a point estimate of the posterior mean \(\mathbb{E}[z \mid \text{history}]\), and the deployed system is a certainty-equivalent adaptive controller re-estimating its own dynamics tens of times per second.

“One Brain, Many Bodies” = Estimation at Scale

Skild Brain’s omni-body claim — driving quadrupeds, humanoids, and manipulators it has not trained on, adapting within seconds to payload shifts or damaged joints — is the adaptation-module loop scaled up.

Why history suffices

If your commanded action \(a_{t-1}\) and the resulting state \(x_t\) disagree with the nominal model, the discrepancy is informative about \(\xi\). The module \(\phi\) learns the inverse map

\[ (\text{surprise in dynamics}) \longmapsto \hat z_t \]

Why more sensing helps

The estimate’s error is controlled by the observability Gramian of the excited dynamics:

  • error is monotone-decreasing in the sensor set (Loewner order)
  • improves as \(O(m^{-1/2})\) with \(m\) independent channels

This is the theoretical reason richer proprioception and force sensing keep appearing on frontier hardware.

Domain Randomization = Stochastic Robust Optimization

Every sim-trained locomotion stack (Unitree-class humanoids, Skild, Tesla sim RL) trains under a distribution over dynamics parameters:

\[ \max_\theta\ \mathbb{E}_{\xi \sim p(\xi)}\; \mathbb{E}_{\pi_\theta,\,\xi}\Bigl[\textstyle\sum_t \gamma^t\, r(x_t, a_t;\, \xi)\Bigr] \]

  • The Bayes-risk (average-case) relaxation of minimax robust control
  • Adaptive curricula that shift \(p(\xi)\) toward the policy’s failure set push this toward the distributionally robust version: \(\ \min_{q:\,\mathrm{KL}(q\Vert p)\le\rho}\ \mathbb{E}_q[\cdot]\)

Robust vs. adaptive — the gap the estimator buys

Conditioning the policy on \(\hat z_t\) converts the robust solution (one controller for all \(\xi\)) into an adaptive one (a \(\xi\)-indexed family). The performance gap between the two is precisely the value of the adaptation module.

Residual dynamics (e.g., ASAP): estimate not \(\xi\) but the sim-to-real discrepancy itself — fit \(\Delta_\eta\) so \(f_{\mathrm{sim}}(x,\, a + \Delta_\eta(x,a)) \approx f_{\mathrm{real}}(x,a)\), then keep training RL in the corrected simulator.

Part III — Reinforcement Learning as Mirror Descent

RL Refresher: Objective and Advantage

Markov decision process: states \(x\), actions \(a\), reward \(r\), discount \(\gamma \in (0,1)\), policy \(\pi_\theta(a \mid x)\).

\[ J(\theta) = \mathbb{E}_{\pi_\theta}\Bigl[\textstyle\sum_{t\ge 0} \gamma^t\, r(x_t, a_t)\Bigr] \]

The three value objects:

\[ V^\pi(x) = \mathbb{E}_\pi\bigl[\text{return} \mid x_0 = x\bigr], \qquad Q^\pi(x,a) = \mathbb{E}_\pi\bigl[\text{return} \mid x_0 = x,\, a_0 = a\bigr] \]

\[ A^\pi(x,a) = Q^\pi(x,a) - V^\pi(x) \quad\text{= "how much better than my average behavior is } a \text{?"} \]

The core tension of policy optimization

The advantage \(A^{\pi_k}\) is only trustworthy near the policy \(\pi_k\) that generated the data. Improve too aggressively and the estimate is garbage; too timidly and learning stalls. Every method on the next slides is a way of formalizing “improve, but stay close.”

PPO: The Locomotion Workhorse

The results that made humanoid locomotion credible (Unitree-class pipelines: 4096 parallel simulated environments, 50 Hz policies) rest on PPO, the practical form of the trust-region problem:

\[ \max_\theta\ \mathbb{E}_{\pi_k}\!\Bigl[\frac{\pi_\theta(a \mid x)}{\pi_k(a \mid x)}\, A^{\pi_k}(x,a)\Bigr] \quad\text{s.t.}\quad \mathbb{E}_{x}\ \mathrm{KL}\bigl(\pi_k(\cdot \mid x)\ \Vert\ \pi_\theta(\cdot \mid x)\bigr) \le \delta \]

  • Objective: importance-weighted advantage — improve where \(A > 0\), retreat where \(A < 0\)
  • Constraint: stay inside a KL ball around the data-generating policy
  • PPO’s clipped ratio \(\min\bigl(\rho\, A,\ \mathrm{clip}(\rho, 1\pm\varepsilon_c)\, A\bigr)\) enforces the proximal region coordinatewise — same principle, no explicit constraint

Next: solving the Lagrangian of this problem in closed form reveals the deeper structure.

The Mirror Descent View

Replace the hard constraint by a KL penalty and solve per state, in closed form:

\[ \pi_{k+1}(\cdot \mid x) = \arg\max_\pi\; \mathbb{E}_{\pi}\bigl[A^{\pi_k}(x, \cdot)\bigr] - \tfrac{1}{\eta}\,\mathrm{KL}\bigl(\pi\ \Vert\ \pi_k(\cdot \mid x)\bigr) \]

\[ \Longrightarrow\qquad \boxed{\ \pi_{k+1}(a \mid x)\ \propto\ \pi_k(a \mid x)\; e^{\eta\, A^{\pi_k}(x,a)}\ } \]

the exponentiated-gradient / multiplicative-weights update — mirror descent on the policy simplex with the KL Bregman divergence.

One update, three famous algorithms

  • Natural policy gradient = its small-step limit (the KL Hessian is the Fisher metric)
  • TRPO/PPO = the same step with the KL region enforced by constraint / clipping
  • AWR & Recap (next) = the same step projected onto a parametric family

Advantage-Weighted Regression: Offline Mirror Descent

We have the ideal update \(\pi_{k+1} \propto \pi_k\, e^{\eta A}\) — but \(\pi_{k+1}\) must live in our parametric family. Project it back by maximum likelihood (an M-projection, like BC):

\[ \min_\theta\ \mathrm{KL}\bigl(\pi_{k+1}\ \Vert\ \pi_\theta\bigr) \;\;=\;\; \min_\theta\ -\,\mathbb{E}_{(x,a)\sim\mathcal D} \Bigl[\, e^{A(x,a)/\beta}\ \log \pi_\theta(a \mid x)\Bigr] + \text{const} \]

Read it as: behavior cloning, where each datapoint is weighted by the exponentiated advantage — clone the good actions hard, the mediocre ones barely.

  • Works offline: \(\mathcal D\) can mix demonstrations, autonomous rollouts, corrections
  • One conservative mirror-descent step from whatever data exists
  • The KL anchor to \(\pi_k\) is what keeps it safe on hardware

BC is the special case \(A \equiv \text{const}\): uniform weights. RL fine-tuning = reweighted imitation.

Recap (\(\pi^*_{0.6}\)): Mirror Descent by Conditioning

Physical Intelligence’s Recap is the frontier instantiation of the AWR step — engineered for a multimodal flow-matching head, where per-sample loss weights are awkward.

Recipe:

  1. Train a value function by temporal-difference / expectile regression on heterogeneous data: demos, autonomous rollouts, teleoperated corrections
  2. Train the flow-matching VLA conditioned on the advantage \(\hat A\): \[ p_\psi\bigl(a \mid h_t,\ \hat A\bigr) \quad \text{trained on all data} \]
  3. At deployment, clamp the conditioning: \[ a \sim p_\psi\bigl(\cdot \mid h_t,\ \hat A = \text{high}\bigr) \]

Conditioning + clamping implements the exponential reweighting without touching the loss — the same KL-proximal improvement, robust to the generative head.

Reported effects

\(\sim\!2\times\) throughput, \(\ge 2\times\) failure reduction on long-horizon tasks (espresso, laundry, box assembly) — the first convincing on-robot RL fine-tuning of a large VLA.

Diffusion Meets RL: Two Mathematical Routes

How do you make a diffusion policy maximize return, not just imitate? Two techniques dominate:

1. Advantage-weighted diffusion (offline). Put the AWR weight inside the score-matching loss:

\[ \min_\psi\ \mathbb{E}_{(x,a)\sim\mathcal D,\, s,\, \epsilon}\; \Bigl[\, e^{A(x,a)/\beta}\; \bigl\|\epsilon - \epsilon_\psi\bigl(\sqrt{\bar\alpha_s}\,a + \sqrt{1-\bar\alpha_s}\,\epsilon,\ s,\ x\bigr)\bigr\|^2 \Bigr] \]

— a KL-regularized policy-iteration step that concentrates probability mass on high-return actions while preserving multimodality.

2. Q-guided sampling (at inference). Bias the reverse process with a learned critic. Targeting \(p_\psi \cdot e^{Q/\beta}\) and applying Bayes’ rule to the score:

\[ \tilde{\epsilon}_\psi(a^s, s, x) = \epsilon_\psi(a^s, s, x) - \kappa_s\, \nabla_{a} Q\bigl(x,\ \hat a^0(a^s)\bigr) \]

where \(\hat a^0(a^s)\) is the Tweedie estimate of the clean action and \(\kappa_s\) the guidance strength — steer toward higher return at runtime, no retraining.

Distillation is KL Projection

Teacher–student steps appear everywhere in these stacks:

\[ \min_\theta\ \mathbb{E}_{x}\ \mathrm{KL}\Bigl(\pi_T\bigl(\cdot \mid x_{\mathrm{priv}}\bigr)\ \Big\Vert\ \pi_\theta\bigl(\cdot \mid x_{\mathrm{obs}}\bigr)\Bigr) \]

— the I-projection of a privileged policy onto a deployable information pattern.

Instance Teacher (privileged) Student (deployable)
RMA phase 2 policy seeing true \(\xi\) policy seeing only history
Sim teachers full-state sim policy vision-based policy
NVIDIA HOVER many whole-body controllers one unified controller
Gemini Robotics on-device cloud VLM on-robot model
Helix S2 \(\to\) S1 7B semantic latent 80M motor policy

Conceptually the same object as the adaptation module: compress privileged information into what the robot can actually observe.

Part IV — Dual-System Architectures as Singular Perturbation

Two Brains, Two Timescales

Figure Helix

  • System 2: 7B-param VLM at 7–9 Hz — emits a semantic latent
  • System 1: 80M visuomotor policy at 200 Hz — consumes it
  • Rate separation \(\varepsilon \approx 1/25\)

The same pattern everywhere

  • NVIDIA GR00T N1: VLM \(\sim\!10\) Hz / diffusion S1 at 120 Hz
  • Gemini Robotics: cloud reasoner + local low-latency decoder
  • Chunked VLAs in general: a chunk is a zero-order-hold slow variable

\[ \underbrace{h_{t+1} = h_t + \varepsilon\, g(h_t, u_t, o_t)}_{\text{slow semantic state}} \qquad\qquad \underbrace{u_{t+1} = f(u_t, h_t, o_t)}_{\text{fast motor state, } f \text{ contracting}} \]

This is exactly the discrete-time singularly perturbed system of classical control theory.

Why “Let the VLM Think While the Reflexes Act” Is a Theorem

Tikhonov’s quasi-steady-state argument, applied to robot brains:

  • The fast policy \(f\) contracts on the timescale of a control tick \(\Rightarrow\) it equilibrates against a slowly varying semantic boundary condition \(h_t\)
  • The slow module may therefore plan on the reduced model (fast dynamics replaced by their equilibrium)
  • Decoupled design — slow dynamic programming + fast tracking — loses only \(O(\varepsilon)\) relative to the intractable joint problem

Two engineering facts this explains

Why chunking works at all: within a chunk the semantic intent is frozen (slow), and the open-loop chunk is stable because low-level tracking (fast) rejects disturbances.

Why raising the S1 rate pays disproportionately: the suboptimality scales with the timescale ratio \(\varepsilon\), not with either rate alone — Figure’s move to 200 Hz shrank \(\varepsilon\) directly, and dexterity gains followed.

Synthesis

Every Method, Its Mathematics, Its Loss

Method (users) Mathematical frame Canonical objective
Behavior cloning (all labs) max likelihood / M-projection \(\min_\psi \mathbb{E}[-\log p_\psi(a\mid o,\ell)]\)
Diffusion policy (Atlas, GR00T) hierarchical VI / score matching \(\mathbb{E}\Vert\epsilon - \epsilon_\psi(a^s, s, h)\Vert^2\)
Flow matching (\(\pi_0\), LBM) normalizing flows / optimal transport \(\mathbb{E}\Vert v_\psi(a^\tau,\tau,h) - (a-\epsilon)\Vert^2\)
Tokenized actions (RT-2, FAST) autoregressive MLE + rate–distortion cross-entropy on DCT tokens
World models (1XWM, GR00T) latent state-space ELBO reconstruction \(-\) KL per step
Rapid motor adaptation (Skild) on-line parameter estimation \(\mathbb{E}\Vert\hat z_t - e(\xi)\Vert^2\)
Domain randomization (sim-RL) stochastic robust optimization \(\max_\theta \mathbb{E}_{\xi}\mathbb{E}_{\pi_\theta}[\sum \gamma^t r]\)
PPO locomotion (Unitree-class) mirror descent / natural gradient KL-constrained surrogate
Recap (\(\pi^*_{0.6}\)) KL-proximal policy iteration \(\mathbb{E}[e^{A/\beta}\log\pi_\theta]\)
Distillation (HOVER, Helix) I-projection \(\min\mathrm{KL}(\pi_T \Vert \pi_\theta)\)
Dual-system S2/S1 (Helix, GR00T) singular perturbation \(O(\varepsilon)\) decoupling loss

Three Observations from the Synthesis

1 — Variational inference won the action head

All competitive decoders are trained as generative models of action chunks — diffusion/flow ELBOs or autoregressive MLE — because demonstrator distributions are multimodal. The remaining competition is over sampling latency: few-step ODE vs. many-step chain vs. token loop.

2 — Estimation, not scale, is what transfers across bodies

Omni-body claims rest on the RMA estimation loop; observability theory (sensor monotonicity, \(O(m^{-1/2})\)) gives the scaling law for why richer proprioception helps.

3 — Mirror descent is the common abstraction of the RL layer

PPO for locomotion, AWR/Recap for VLA fine-tuning, distillation’s KL projections — all KL-proximal iterations. The frontier trend: moving the mirror-descent step from simulation onto the physical robot, with human corrections as off-policy data.

Open Challenges at the Frontier

  • Inference latency for high-frequency control. Iterative denoising vs. 100–200 Hz balancing: consistency models, latent diffusion, and few-step flow matching are the active fronts

  • Offline-to-online fine-tuning. The score-matching objective does not directly align with expected return; advantage weighting and Q-guidance bridge the gap without collapsing multimodality — but on-robot RL remains sample-limited

  • Estimation layer \(\not\leftrightarrow\) generative layer. No public stack conditions the VLA action head on the adaptation latent \(\hat z_t\) — though it conditions locomotion policies, and the belief-embedding formalism says they are the same object

  • Out-of-distribution robustness. Generative heads capture the support of the demos perfectly, then extrapolate unpredictably; blending robust control priors with generative action heads is open

  • Certified on-robot policy iteration. Recap is a single conservative mirror-descent step; trust-region methods with certified step sizes would make on-robot iteration routine rather than heroic

Takeaways

  1. One architecture: VLM encoder \(\;+\;\) history-based estimator \(\;+\;\) generative action-chunk head

  2. One training principle for the head: maximize a (variational) likelihood — BC is MLE, diffusion is the ELBO of a noising hierarchy, flow matching is its optimal-transport geodesic twin

  3. One estimation principle for transfer: amortized posterior-mean regression over embodiment parameters — adaptive beats robust by exactly the estimation gap

  4. One RL principle: KL-proximal (mirror-descent) policy iteration — PPO, AWR, and advantage conditioning are the same update in three costumes

  5. One systems principle: timescale separation with \(O(\varepsilon)\) guarantees — chunking and dual-system designs are singular perturbation, engineered deliberately

The vocabulary is new; the mathematics is classical. Knowing the six ideas, you can read any humanoid lab’s technical report — and see exactly which knob they turned.