Gradient-based planning for world fashions at longer horizons

By Michael Psenka, Mike Rabbat, Aditi Krishnapriyan, Yann LeCun, Amir Bar

GRASP is a brand new gradient-based planner for discovered dynamics (a “world mannequin”) that makes long-horizon planning sensible by (1) lifting the trajectory into digital states so optimization is parallel throughout time, (2) including stochasticity on to the state iterates for exploration, and (3) reshaping gradients so actions get clear indicators whereas we keep away from brittle “state-input” gradients by means of high-dimensional imaginative and prescient fashions.

Giant, discovered world fashions have gotten more and more succesful. They will predict lengthy sequences of future observations in high-dimensional visible areas and generalize throughout duties in ways in which have been troublesome to think about a couple of years in the past. As these fashions scale, they begin to look much less like task-specific predictors and extra like general-purpose simulators.

However having a strong predictive mannequin shouldn’t be the identical as having the ability to use it successfully for management/studying/planning. In apply, long-horizon planning with trendy world fashions stays fragile: optimization turns into ill-conditioned, non-greedy construction creates unhealthy native minima, and high-dimensional latent areas introduce refined failure modes.

On this weblog publish, I describe the issues that motivated this mission and our strategy to handle them: why planning with trendy world fashions could be surprisingly fragile, why lengthy horizons are the true stress check, and what we modified to make gradient-based planning rather more strong.

This weblog publish discusses work accomplished with Mike Rabbat, Aditi Krishnapriyan, Yann LeCun, and Amir Bar (* denotes equal advisorship), the place we suggest GRASP.

What’s a world mannequin?

Today, the time period “world mannequin” is kind of overloaded, and relying on the context can both imply an express dynamics mannequin or some implicit, dependable inner state {that a} generative mannequin depends on (e.g. when an LLM generates chess strikes, whether or not there’s some inner illustration of the board). We give our unfastened working definition under.

Suppose you are taking actions $a_t in mathcal{A}$ and observe states $s_t in mathcal{S}$ (pictures, latent vectors, proprioception). A world mannequin is a discovered mannequin that, given the present state and a sequence of future actions, predicts what’s going to occur subsequent. Formally, it defines a predictive distribution on a sequence of noticed states $s_{t-h:t}$ and present motion $a_t$ :

$[P_theta(s_{t+1} mid s_{t-h:t},; a_t)]$

that approximates the setting’s true conditional $P(s_{t+1} mid s_{t-h:t},; a_t)$ . For this weblog publish, we’ll assume a Markovian mannequin $P(s_{t+1} mid s_{t-h:t},; a_t)$ for simplicity (all outcomes right here could be prolonged to the extra basic case), and when the mannequin is deterministic it reduces to a map over states:

$[s_{t+1} = F_theta(s_t, a_t).]$

In apply the state $s_t$ is commonly a discovered latent illustration (e.g., encoded from pixels), so the mannequin operates in a (theoretically) compact, differentiable house. The important thing level is {that a} world mannequin provides you a differentiable simulator; you possibly can roll it ahead beneath hypothetical motion sequences and backpropagate by means of the predictions.

Planning: selecting actions by optimizing by means of the mannequin

Given a begin $s_0$ and a aim $g$ , the only planner chooses an motion sequence $mathbf{a}=(a_0,dots,a_{T-1})$ by rolling out the mannequin and minimizing terminal error:

$[min_{mathbf{a}} ; | s_T(mathbf{a}) - g |_2^2, quad text{where } s_T(mathbf{a}) = mathcal{F}_{theta}^{T}(s_0,mathbf{a}).]$

Right here we use $mathcal{F}^T$ as shorthand for the total rollout by means of the world mannequin (dependence on mannequin parameters $theta$ is implicit):

$[mathcal{F}_{theta}^{T}(s_0, mathbf{a}) = F_theta(F_theta(cdots F_theta(s_0, a_0), cdots, a_{T-2}), a_{T-1}).]$

In brief horizons and low-dimensional techniques, this may work moderately nicely. However as horizons develop and fashions turn into bigger and extra expressive, its weaknesses turn into amplified.

So why doesn’t this simply work at scale?

Why long-horizon planning is difficult (even when every thing is differentiable)

There are two separate ache factors for the extra basic world mannequin, plus a 3rd that’s particular to discovered, deep learning-based fashions.

1) Lengthy-horizon rollouts create deep, ill-conditioned computation graphs

These conversant in backprop by means of time (BPTT) could discover that we’re differentiating by means of a mannequin utilized to itself repeatedly, which is able to result in the exploding/vanishing gradients drawback. Specifically, if we take derivatives (notice we’re differentiating vector-valued capabilities, leading to Jacobians that we denote with $D_x (cdots)$ ) with respect to earlier actions (e.g. $a_0$ ):

$[D_{a_0} mathcal{F}_{theta}^{T}(s_0, mathbf{a}) = Bigl(prod_{t=1}^T D_s F_theta(s_t, a_t)Bigr) D_{a_0}F_theta(s_0, a_0).]$

We see that the Jacobian’s conditioning scales exponentially with time $T$ :

$[sigma_{text{max/min}}(D_{a_0}mathcal{F}_{theta}^{T}) sim sigma_{text{max/min}}(D_s F_theta)^{T-1},]$

resulting in exploding or vanishing gradients.

2) The panorama is non-greedy and filled with traps

At brief horizons, the grasping resolution, the place we transfer straight towards the aim at each step, is commonly ok. Should you solely have to plan a couple of steps forward, the optimum trajectory often doesn’t deviate a lot from “head towards $g$ ” at every step.

As horizons develop, two issues occur. First, longer duties usually tend to require non-greedy habits: going round a wall, repositioning earlier than pushing, backing as much as take a greater path. And as horizons develop, extra of those non-greedy steps are sometimes wanted. Second, the optimization house itself scales with horizon: $mathrm{dim}(mathcal{A} times cdots times mathcal{A}) = Tmathrm{dim}(mathcal{A})$ , additional increasing the house of native minima for the optimization drawback.

Loss landscape — *Distance to aim alongside the optimum path is non-monotonic, and the ensuing loss panorama could be tough.*

An extended-horizon repair: lifting the dynamics constraint

Suppose we deal with the dynamics constraint $s_{t+1} = F_{theta}(s_t, a_t)$ as a tender constraint, and we as an alternative optimize the next penalty perform over each actions $(a_0,ldots,a_{T-1})$ and states $(s_0,ldots,s_T)$ :

$[min_{mathbf{s},mathbf{a}} mathcal{L}(mathbf{s}, mathbf{a}) = sum_{t=0}^{T-1} big|F_theta(s_t,a_t) - s_{t+1}big|_2^2, quad text{with } s_0 text{ fixed and } s_T=g.]$

That is additionally generally referred to as collocation in planning/robotics literature. Word the lifted formulation shares the identical world minimizers as the unique rollout goal (each are zero precisely when the trajectory is dynamically possible). However the optimization landscapes are very completely different, and we get two instant advantages:

Every world mannequin analysis $F_{theta}(s_t,a_t)$ relies upon solely on native variables, so all $T$ phrases could be computed in parallel throughout time, leading to an enormous speed-up for longer horizons, and
You not backpropagate by means of a single deep $T$ -step composition to get a studying sign, for the reason that earlier product of Jacobians now splits right into a sum, e.g.:

$[D_{a_0} mathcal{L} = 2(F_theta(s_0, a_0) - s_1).]$

With the ability to optimize states straight additionally helps with exploration, as we are able to quickly navigate by means of unphysical domains to seek out the optimum plan:

Collocation planning in BallNav — *Collocation-based planning permits us to straight perturb states and discover midpoints extra successfully.*

Nonetheless, lunch is rarely free. And certainly, particularly for deep learning-based world fashions, there’s a essential situation that makes the above optimization fairly troublesome in apply.

A problem for deep learning-based world fashions: sensitivity of state-input gradients

The tl;dr of this part is: straight optimizing states by means of a deep learning-based $F_{theta}$ is extremely brittle, à la adversarial robustness. Even should you prepare your world mannequin in a lower-dimensional state house, the coaching course of for the world mannequin makes unseen state landscapes very sharp, whether or not it’s an unseen state itself or just a standard/orthogonal route to the information manifold.

Adversarial robustness and the “dimpled manifold” mannequin

Adversarial robustness initially checked out classification fashions $f_theta : mathbb{R}^{wtimes h times c} to mathbb{R}^K$ , and confirmed that by following the gradient of a specific logit $nabla f_theta^k$ from a base picture $x$ (not of sophistication $k$ ), you didn’t have to maneuver far alongside $x' = x + epsilonnabla f_theta^k$ to make $f_theta$ classify $x'$ as $k$ (Szegedy et al., 2014; Goodfellow et al., 2015):

Adversarial example — *Depiction of the traditional instance from (Goodfellow et al., 2015).*

Later work has painted a geometrical image for what’s occurring: for information close to a low-dimensional manifold $mathcal{M}$ , the coaching course of controls habits in tangential instructions, however doesn’t regularize habits in orthogonal instructions, thus resulting in delicate habits (Stutz et al., 2019). One other approach acknowledged: $f_theta$ has an affordable Lipschitz fixed when contemplating solely tangential instructions to the information manifold $mathcal{M}$ , however can have very excessive Lipschitz constants in regular instructions. In truth, it typically advantages the mannequin to be sharper in these regular instructions, so it will possibly match extra sophisticated capabilities extra exactly.

Adversarial perturbations leave the data manifold

Because of this, such adversarial examples are extremely widespread even for a single given mannequin. Additional, this isn’t simply a pc imaginative and prescient phenomenon; adversarial examples additionally seem in LLMs (Wallace et al., 2019) and in RL (Gleave et al., 2019).

Whereas there are strategies to coach for extra adversarially strong fashions, there’s a recognized trade-off between mannequin efficiency and adversarial robustness (Tsipras et al., 2019): particularly within the presence of many weakly-correlated variables, the mannequin should be sharper to realize larger efficiency. Certainly, most trendy coaching algorithms, whether or not in laptop imaginative and prescient or LLMs, don’t prepare adversarial robustness out. Thus, at the very least till deep studying sees a serious regime change, it is a drawback we’re caught with.

Why is adversarial robustness a problem for world mannequin planning?

Contemplate a single element of the dynamics loss we’re optimizing within the lifted state strategy:

$[min_{s_t, a_t, s_{t+1}} |F_theta(s_t, a_t) - s_{t+1}|_2^2]$

Let’s additional give attention to simply the bottom state:

$[min_{s_t} |F_theta(s_t, a_t) - s_{t+1}|_2^2.]$

Since world fashions are sometimes skilled on state/motion trajectories $(s_1, a_1, s_2, a_2, ldots)$ , the state-data manifold for $F_{theta}$ has dimensionality bounded by the motion house:

$[mathrm{dim}(mathcal{M}_s) le mathrm{dim}(mathcal{A}) + 1 + mathrm{dim}(mathcal{R}),]$

the place $mathcal{R}$ is a few optionally available house of augmentations (e.g. translations/rotations). Thus, we are able to sometimes count on $mathrm{dim}(mathcal{M}_s)$ to be a lot decrease than $mathrm{dim}(mathcal{S})$ , and thus: it is vitally straightforward to seek out adversarial examples that hack any state to another desired state.

Because of this, the dynamics optimization

$[sum_{t=0}^{T-1} big|F_theta(s_t,a_t) - s_{t+1}big|_2^2]$

feels extremely “sticky,” as the bottom factors $s_t$ can simply trick $F_{theta}$ into pondering it’s already made its native aim.¹

1. This adversarial robustness situation, whereas notably unhealthy for lifted-state approaches, shouldn’t be distinctive to them. Even for serial optimization strategies that optimize by means of the total rollout map $mathcal{F}^T$ , it’s potential to get into unseen states, the place it is vitally straightforward to have a standard element fed into the delicate regular parts of $D_s F_{theta}$ . The motion Jacobian’s chain rule growth is

$[Bigl(prod_{t=1}^T D_s F_theta(s_t, a_t)Bigr) D_{a_0}F_theta(s_0, a_0).]$

See what occurs if any stage of the product has any element regular to the information manifold. ↩

Our repair

That is the place our new planner GRASP is available in. The principle remark: whereas $D_s F_{theta}$ is untrustworthy and adversarial, the motion house is often low-dimensional and exhaustively skilled, so $D_a F_{theta}$ is definitely affordable to optimize by means of and doesn’t undergo from the adversarial robustness situation!

Network diagram showing high-dim state vs low-dim action — *The motion enter is often lower-dimensional and densely skilled (the mannequin has seen each motion route), so motion gradients are a lot better behaved.*

At its core, GRASP builds a first-order lifted state / collocation-based planner that’s solely depending on motion Jacobians by means of the world mannequin. We thus exploit the differentiability of discovered world fashions $F_{theta}$ , whereas not falling sufferer to the inherent sensitivity of the state Jacobians $D_s F_{theta}$ .

GRASP: Gradient RelAxed Stochastic Planner

As famous earlier than, we begin with the collocation planning goal, the place we elevate the states and calm down dynamics right into a penalty:

$[min_{mathbf{s},mathbf{a}} mathcal{L}(mathbf{s}, mathbf{a}) = sum_{t=0}^{T-1} big|F_theta(s_t,a_t) - s_{t+1}big|_2^2, quad text{with } s_0 text{ fixed and } s_T=g.]$

We then make two key additions.

Ingredient 1: Exploration by noising the state iterates

Even with a smoother goal, planning is nonconvex. We introduce exploration by injecting Gaussian noise into the digital state updates throughout optimization.

A easy model:

$[s_t leftarrow s_t - eta_s nabla_{s_t}mathcal{L} + sigma_{text{state}} xi, qquad xisimmathcal{N}(0,I).]$

Actions are nonetheless up to date by non-stochastic descent:

$[a_t leftarrow a_t - eta_a nabla_{a_t}mathcal{L}.]$

The state noise helps you “hop” between basins within the lifted house, whereas the actions stay guided by gradients. We discovered that particularly noising states right here (versus actions) finds an excellent steadiness of exploration and the power to seek out sharper minima.²

2. As a result of we solely noise the states (and never the actions), the corresponding dynamics usually are not actually Langevin dynamics. ↩

Ingredient 2: Reshape gradients: cease brittle state-input gradients, hold motion gradients

As mentioned, the delicate pathway is the gradient that flows into the state enter of the world mannequin, $D_s F_{theta}$ . Probably the most simple approach to do that initially is to only cease state gradients into $F_{theta}$ straight:

Let $bar{s}_t$ be the identical worth as $s_t$ , however with gradients stopped.

Outline the stop-gradient dynamics loss:

$[mathcal{L}_{text{dyn}}^{text{sg}}(mathbf{s},mathbf{a}) = sum_{t=0}^{T-1} big|F_theta(bar{s}_t, a_t) - s_{t+1}big|_2^2.]$

This alone doesn’t work. Discover now states solely observe the earlier state’s step, with out something forcing the bottom states to chase the following ones. Because of this, there are trivial minima for simply stopping on the origin, then just for the ultimate motion attempting to get to the aim in a single step.

Dense aim shaping

We are able to view the above situation because the aim’s sign being lower off solely from earlier states. One strategy to repair that is to easily add a dense aim time period all through prediction:

$[mathcal{L}_{text{goal}}^{text{sg}}(mathbf{s},mathbf{a}) = sum_{t=0}^{T-1} big|F_theta(bar{s}_t, a_t) - gbig|_2^2.]$

In regular settings this is able to over-bias in direction of the grasping resolution of straight chasing the aim, however that is balanced in our setting by the stop-gradient dynamics loss’s bias in direction of possible dynamics. The ultimate goal is then as follows:

$[mathcal{L}(mathbf{s},mathbf{a}) = mathcal{L}_{text{dyn}}^{text{sg}}(mathbf{s},mathbf{a}) + gamma , mathcal{L}_{text{goal}}^{text{sg}}(mathbf{s},mathbf{a}).]$

The result’s a planning optimization goal that doesn’t have dependence on state gradients.

Periodic “sync”: briefly return to true rollout gradients

The lifted stop-gradient goal is nice for quick, guided exploration, but it surely’s nonetheless an approximation of the unique serial rollout goal.

So each $K_{text{sync}}$ iterations, GRASP does a brief refinement part:

Roll out from $s_0$ utilizing present actions $mathbf{a}$ , and take a couple of small gradient steps on the unique serial loss:

$[mathbf{a} leftarrow mathbf{a} - eta_{text{sync}},nabla_{mathbf{a}},|s_T(mathbf{a})-g|_2^2.]$

The lifted-state optimization nonetheless offers the core of the optimization, whereas this refinement step provides some help to maintain states and actions grounded in direction of actual trajectories. This refinement step can after all get replaced with a serial planner of your alternative (e.g. CEM); the core thought is to nonetheless get a few of the advantage of the full-path synchronization of serial planners, whereas nonetheless largely utilizing the advantages of the lifted-state planning.

How GRASP addresses long-range planning

Collocation-based planners provide a pure repair for long-horizon planning, however this optimization is kind of troublesome by means of trendy world fashions resulting from adversarial robustness points. GRASP proposes a easy resolution for a smoother collocation-based planner, alongside secure stochasticity for exploration. Because of this, longer-horizon planning finally ends up not solely succeeding extra, but in addition discovering such successes quicker:

Push-T planning demo — *Push-T demo: longer-horizon planning with GRASP.*

Horizon	CEM	GD	LatCo	GRASP
H=40	61.4% / 35.3s	51.0% / 18.0s	15.0% / 598.0s	59.0% / 8.5s
H=50	30.2% / 96.2s	37.6% / 76.3s	4.2% / 1114.7s	43.4% / 15.2s
H=60	7.2% / 83.1s	16.4% / 146.5s	2.0% / 231.5s	26.2% / 49.1s
H=70	7.8% / 156.1s	12.0% / 103.1s	0.0% / —	16.0% / 79.9s
H=80	2.8% / 132.2s	6.4% / 161.3s	0.0% / —	10.4% / 58.9s

Push-T outcomes. Success charge (%) / median time to success. Daring = finest in row. Word the median success time will bias larger with larger success charge; GRASP manages to be quicker regardless of larger success charge.

What’s subsequent?

There may be nonetheless loads of work to be accomplished for contemporary world mannequin planners. We wish to exploit the gradient construction of discovered world fashions, and collocation (lifted-state optimization) is a pure strategy for long-horizon planning, but it surely’s essential to know typical gradient construction right here: easy and informative motion gradients and brittle state gradients. We view GRASP as an preliminary iteration for such planners.

Extension to diffusion-based world fashions (deeper latent timesteps could be seen as smoothed variations of the world mannequin itself), extra subtle optimizers and noising methods, and integrating GRASP into both a closed-loop system or RL coverage studying for adaptive long-horizon planning are all pure and fascinating subsequent steps.

I do genuinely assume it’s an thrilling time to be engaged on world mannequin planners. It’s a humorous candy spot the place the background literature (planning and management general) is extremely mature and well-developed, however the present setting (pure planning optimization over trendy, large-scale world fashions) continues to be closely underexplored. However, as soon as we determine all the precise concepts, world mannequin planners will possible turn into as commonplace as RL.

For extra particulars, learn the full paper or go to the mission web site.

Quotation

@article{psenka2026grasp,
  title={Parallel Stochastic Gradient-Based mostly Planning for World Fashions},
  writer={Michael Psenka and Michael Rabbat and Aditi Krishnapriyan and Yann LeCun and Amir Bar},
  12 months={2026},
  eprint={2602.00475},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2602.00475}
}

This text was initially printed on the BAIR weblog, and seems right here with the authors’ permission.

BAIR Weblog
is the official weblog of the Berkeley Synthetic Intelligence Analysis (BAIR) Lab.