Candidate solutions and research bets

Why this chapter exists

Chapter 17 listed what is unsolved. This chapter lists what is being tried. For each problem it names the candidate solutions visible in the source base, who is trying them, and an assessment of how far along each is. The assessment uses three levels. Established means in shipped models or standard practice, with more than one group reporting it works. Promising means demonstrated by more than one group or on real robots, but not yet standard. Early means one or two papers, often only in simulation.

"Promising" is a judgement. Each assessment is dated October 2026, and the chapter expects to be revised.

After this chapter you will be able to:

• name, for each open problem, at least one candidate solution and say how mature it is;
• say which families of figure 1.1 each solution would change, using figure 18.2;
• explain with figure 18.3 why spending more compute at test time helps only as much as the verifier allows.

Where the solutions land

Most solutions do not fix a problem in one family; they shift where families sit relative to each other. Video-only-in-training makes world-action models cost what VLAs cost at inference. Latent prediction pulls world-action models towards chapter 7. Explicit 3D state pulls them towards chapter 9. Figure 18.2 draws those moves on the family grid.

Action alignment and factorisation

Action alignment. The WAM survey's two recommendations are the main line. The first is staged training, which aligns a pretrained video model's predicted motion to executable actions on embodied data. The second is prior preservation through frozen parameters, alignment layers and residual action heads (Lu et al. 2026, arXiv 2609.16074). Both are promising. Latent action interfaces with small per-robot decoders (chapter 7) are a second promising route (Ye et al. 2024, arXiv 2410.11758; Routray et al. 2025, arXiv 2511.07732), and using an inverse dynamics model as a reward for executability is early (Wang et al. 2026, arXiv 2603.17808).

Factorisation. Training with futures and acting without them is the most visible bet. Fast-WAM keeps video co-training and drops future generation at test time (Yuan et al. 2026, arXiv 2603.16666), and LingBot-VLA 2.0 reaches the same design from the VLA side (Wu et al. 2026, arXiv 2607.06403). Privileged Foresight Distillation sharpens the idea. It treats what the future teaches as a correction to the action, measured by a teacher that sees the future, and distils it into a small adapter on a student that never generates video (Fang et al. 2026, arXiv 2604.25859). Predicting representations instead of pixels, as JEPA-WAM and LaWAM do, is the other promising bet (Lin et al. 2026, arXiv 2608.09381; Chen et al. 2026, arXiv 2606.15768).

Consistency, memory and neural simulators

Multi-view consistency. Explicit 3D state (chapter 9) is the structural answer, and is promising (Huang et al. 2026, arXiv 2601.03782; Lu et al. 2025, arXiv 2508.17600). A July 2026 paper turns chapter 17's consistency check into a selection rule. It samples several rollouts from a world-action model and keeps the one whose predicted futures agree best across views, measured by depth reprojection with a frozen geometry model (Zhao et al. 2026, arXiv 2607.17454). With eight samples it raised RoboCasa success from 66.3 to 68.4 percent with Cosmos Policy, and from 80.8 to 82.5 percent with another backbone. A gate that only samples when the first rollout looks inconsistent kept about three-quarters of that gain while sampling extra at about a quarter of decisions (Zhao et al. 2026, arXiv 2607.17454). Those modest gains are what figure 18.3 predicts for an imperfect verifier.

Memory. MEM's video-plus-text memory is promising, published by one lab and built into π0.7 (Torne et al. 2026, arXiv 2603.03596; Physical Intelligence et al. 2026, arXiv 2604.15483). The alternative is an agentic harness: a reasoning model that keeps state outside the policy, as Gemini Robotics ER 2 does (Google DeepMind 2026).

Neural simulators. Closing the loop, so the world model is retrained on the policy's own rollouts, is promising (Liu et al. 2026, arXiv 2602.06508; Wang et al. 2026, arXiv 2603.08546). Dense teacher supervision instead of sparse reward (Yang et al. 2026, arXiv 2608.22364) and unifying search with value learning (Cheng et al. 2026, arXiv 2605.26282) are early.

Spending compute at test time

A general bet that cuts across problems 3, 5 and 6 is to spend more computation at decision time. Sample several candidate plans, let a verifier check them, and execute one the verifier accepts. The verifier can be a world model imagining each plan's outcome, a geometric consistency check, or a reasoning model judging the plan. If one candidate succeeds with probability $p$, the verifier accepts good plans with probability $a$ and bad ones with probability $b$, and candidates are checked one at a time up to $N$, then the chance of success is

Read it from the left. $q$ is the chance that any one candidate is accepted. $(1-q)^N$ is the chance that none of the $N$ is, in which case the robot executes the last, rejected one. The first fraction is the verifier's precision, the chance that an accepted plan is good. As $N$ grows, $P_N$ approaches that precision and never exceeds it.

With a policy that succeeds half the time and a verifier that accepts 90 percent of good plans and 20 percent of bad ones, the ceiling is about 82 percent, reached within about eight candidates. Cutting the verifier's false acceptances from 20 to 5 percent raises the ceiling to about 95 percent. The lesson for research bets is that test-time compute is worth as much as the verifier is good, which is why verification is becoming a research problem of its own. FFDC, from May 2026, uses a learned verifier the other way round: to decide when an imagined plan has stopped matching reality, so the robot replans only when needed. It reports up to 74 percent fewer planning calls on a real humanoid, with higher success (Wang et al. 2026, arXiv 2605.06222).

Latency, data, evaluation, safety, embodiment

Latency has the most mature answers. Asynchronous execution and real-time chunking, few-step denoising and parallel decoding are established (Black et al. 2025, arXiv 2506.07339; Kim et al. 2025, arXiv 2502.19645). Right-sized, quantised edge deployment is promising (Zhou et al. 2026, arXiv 2604.24447; NVIDIA 2026), and adaptive replanning is early (Wang et al. 2026, arXiv 2605.06222).

Data. Egocentric human video at scale is promising and heavily funded (Zheng et al. 2026, arXiv 2602.16710; Gao et al. 2026, arXiv 2602.06949; Dyna Robotics 2026; Generalist AI 2026). Synthetic rewriting is promising, with its exchange rate against real data unsettled (Jang et al. 2025, arXiv 2505.12705; NVIDIA et al. 2025, arXiv 2503.14492). Learning from failures through labelled prompts is early (Physical Intelligence et al. 2026, arXiv 2604.15483).

Evaluation. Distributed double-blind real-robot comparison and blind A/B trials with statistical confidence are promising (Atreya et al. 2025, arXiv 2506.18123; TRI LBM Team et al. 2025, arXiv 2507.05331). Perturbation and safety-aware benchmarks are early (Fei et al. 2025, arXiv 2510.13626; Xu et al. 2026, arXiv 2604.18000; Rasouli et al. 2026, arXiv 2604.21192). Readiness frameworks are early too, with one detailed instance so far (Kube et al. 2026, arXiv 2603.06749).

Safety. The only established answer is architectural: a certified envelope around the learned policy (chapter 15). Reasoning models as runtime monitors, judging whether a step succeeded or a plan is unsafe, are early (Google DeepMind 2026). Physical fallbacks for balancing robots are an open question (Rodney Brooks 2026).

Embodiment. Latent tokens to a learned whole-body controller is promising, shipped in GR00T N1.7 (NVIDIA 2026; Luo et al. 2025, arXiv 2511.07820). Cross-embodiment pretraining is promising (Open X-Embodiment Collaboration et al. 2023, arXiv 2310.08864; Octo Model Team et al. 2024, arXiv 2405.12213). Embodiment-agnostic action spaces, such as PointWorld's 3D point flows, are early (Huang et al. 2026, arXiv 2601.03782).

In code

The crate's verifier has two numbers, and the procedure is a short loop.

/// A verifier that checks candidate plans before one is executed.
#[derive(Clone, Copy, Debug)]
pub struct Verifier {
    /// Chance it accepts a plan that would succeed.
    pub accept_good: f64,
    /// Chance it accepts a plan that would fail.
    pub accept_bad: f64,
}

/// The policy's chance that any one sampled plan succeeds, and the verifier that screens them.
/// Candidates are checked one at a time; the first accepted plan is executed. If none of `n` is
/// accepted, the last one is executed anyway.
#[derive(Clone, Copy, Debug)]
pub struct Setup {
    pub p_good: f64,
    pub verifier: Verifier,
}

impl Setup {
    /// Chance that a single candidate is accepted.
    pub fn accept_rate(&self) -> f64 {
        self.p_good * self.verifier.accept_good + (1.0 - self.p_good) * self.verifier.accept_bad
    }
}

/// A small deterministic generator, so simulated runs can be compared with the exact answers.
pub struct Rng(u64);
impl Rng {
    pub fn new(seed: u64) -> Rng {
        Rng(seed)
    }
    pub fn next_f64(&mut self) -> f64 {
        self.0 = self.0.wrapping_add(0x9E37_79B9_7F4A_7C15);
        let mut z = self.0;
        z = (z ^ (z >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
        z = (z ^ (z >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
        ((z ^ (z >> 31)) >> 11) as f64 / (1u64 << 53) as f64
    }
}

/// Run the procedure once: returns (success, candidates examined).
pub fn run_once(s: &Setup, n: usize, rng: &mut Rng) -> (bool, usize) {
    for k in 1..=n {
        let good = rng.next_f64() < s.p_good;
        let accept = rng.next_f64() < if good { s.verifier.accept_good } else { s.verifier.accept_bad };
        if accept || k == n {
            return (good, k);
        }
    }
    unreachable!("n must be at least 1")
}

Exercises

The stubs are in rust/ch18-research-bets/src/exercises.rs.

cargo test -p ch18-research-bets --test exercises ex18_1

cargo test -p ch18-research-bets --test exercises ex18_2

cargo test -p ch18-research-bets --test exercises ex18_3

Further reading

• The WAM survey's section on open challenges and future directions (Lu et al. 2026, arXiv 2609.16074).
• The world-model-to-WAM tutorial, for a compact map of the approaches (Zhang et al. 2026, arXiv 2607.00836).
• Test-time scaling for WAMs, for verification by geometric consistency (Zhao et al. 2026, arXiv 2607.17454).
• When to Trust Imagination, for verifying a plan against reality as it executes (Wang et al. 2026, arXiv 2605.06222).
• Privileged Foresight Distillation, for keeping the benefit of futures without generating them (Fang et al. 2026, arXiv 2604.25859).
• LeJEPA, for the representation-learning bet (Balestriero et al. 2025, arXiv 2511.08544).