The Inference Report

The CoRL 2026 RoboEval challenge has eight bimanual tasks, and it ranks a policy not only on whether it succeeds but on how: every stage of the task, done safely, efficiently and with both arms in coordination, scored against human demonstrations. My solution has two parts. The generic part, carried over from my LeHome challenge solution, is off-policy RL on a $π_{0.5}$-based policy that is also its own value function. It is trained with AWR and RECAP on advantages of the exact official score, in a loop where every rollout is at the same time a Thompson-sampling trial of about 20 inference parameters. Out of the box, it took the policy to 0.85 RDS on held-out seeds in two days (and it was still very far from convergence). The competition-specific part adds data that local exploration cannot find: generated demonstrations that reach the full reward, and recovery episodes that start from the policy's own failures. The submitted policy reaches 0.990 RDS on held-out seeds and 0.987 on the official seeds.

Recent studies of embodied agents based on frontier multimodal large language models (MLLMs) develop harness engineering around the MLLM, such as harnesses with post-trained policies and recursive self-improvement loops. Yet whether these additions exceed the native capability of the MLLM remains unexplored. We introduce SimplEvo to explore how far an MLLM can improve under direct control when it retains its execution experience in self-written text memories. It requires no weight updates, post-trained policies, or privileged information. The MLLM writes a coordinate-free task memory once from a single demonstration. Under a strict train-test protocol, it then creates a bounded working memory on training episodes only and refines it after each episode with lessons on procedure, completion, and recovery. Both memories stay fixed at test time. On EBench, SimplEvo lifts the success rate (SR) of GPT-6-Astra under direct control from 46.7% to 81.7%, and tops the RoboDojo leaderboard with 69.63% SR and a score of 75.32. Compared with memories that also store past visual frames, SimplEvo reaches a similar SR with text memory alone while using fewer tokens. With its memories fixed, SimplEvo also generalizes to unseen layouts and randomized scene configurations. Moreover, the working memory transfers across models: written by GPT-6-Astra, it can be reused as is by weaker MLLMs and improve their performance without any adaptation of their own.

World models learn internal simulators of environment dynamics for prediction, planning, and control, yet learning them from limited, noisy, and partial observations remains challenging. While much progress has focused on pixel-based observations, many physical systems instead provide vector-valued measurements that reveal only part of the underlying state. We introduce K-PWM, a probabilistic world model for prediction and control under partial observation. K-PWM couples Koopman-bilinear latent dynamics with a nonlinear observation decoder and performs probabilistic latent-state inference through the learned model via extended Kalman filtering and smoothing, without a separately trained inference network. Its bilinear structure enables exact propagation of the maintained Gaussian belief through the latent transition and efficient rollouts for uncertainty-aware model predictive path integral (MPPI) control. For training K-PWM, we present a generalized expectation conditional maximization (GECM) procedure with closed-form updates for the generators, noise covariances, and latent prior, complemented by gradient-based decoder learning. A Koopman-consistent initialization selects the latent dimension from data and initializes the action-dependent dynamics. With limited training data, K-PWM achieves lower long-horizon prediction error than the evaluated baselines on partially observed Gymnasium MuJoCo systems, while substantially reducing training time relative to the strongest prediction baseline. On high-dimensional MuJoCo Menagerie robots, it achieves competitive tracking accuracy with low per-step computation times, demonstrating potential for real-time deployment. These results support structured probabilistic dynamics as a data- and compute-efficient approach to world modeling and control under partial observation.

Humanoid loco-manipulation requires continuous coordination of high-dimensional whole-body motion to realize task objectives. While vision-language-action models and world action models transfer semantic and visual-dynamics priors to robot control, these priors do not necessarily capture the structure of robot motion. We introduce Kinematic Action Models (KAMs), a paradigm that treats whole-body motion as a direct pretraining target and uses a pretrained motion generator as the primary action-generation backbone. KAMs ground learned temporal and cross-body coordination priors through task conditioning, visual observations, and execution feedback. We present Kimodo-Policy as an initial instantiation, combining a frozen Kimodo text-to-motion backbone with lightweight ControlNet-based visual conditioning. A root-body-hand hierarchy coordinates locomotion and manipulation, while motion-history conditioning and repeated action-chunk prediction bridge motion synthesis and closed-loop control. Following large-scale loco-manipulation pretraining, the policy achieves the highest overall success among compared policies on multi-task HumanoidArena (72.14%), competitive performance on SIMPLE, and 60% mean success across eight real-world tasks, with 678M parameters (92M trainable). Ablations support the contribution of the pretrained motion backbone. Together, these results provide initial evidence for motion generation as a complementary foundation for humanoid action learning.

Hybrid quantum reinforcement learning has been explored for robotic decision-making, but evaluations often emphasize task success, return, or convergence, leaving unresolved whether quantum circuit outputs influence constrained actions and remain traceable through robot execution. We present \texttt{Qubit-to-Motion}, an evidence-centered framework for support-constrained sequencing and simulated robotic placement. Four Pauli-$Z$ readouts provide the sole learned input to the actor and critic, while a support mask enforces admissible object selection. Each decision is recorded and linked via step indices and SHA-256 fingerprints to a sealed route, deterministic Cartesian and inverse-kinematics waypoints, prescribed joint motion, and MuJoCo contact replay. A frozen checkpoint is evaluated across five deterministic 30-object scene variants. All 150 placements satisfy support, containment, and settling criteria. A fixed-state zero-readout intervention changes $90.0\pm0.0\%$ of decisions, demonstrating dependence on the joint circuit representation. Mean macro travel is $34.537\pm0.206$~m, inference requires $8.60\pm1.27$~ms per route, and replay achieves $8.764\pm0.220~μ$m mean final position error with 0.356~mm maximum interpenetration. These results establish reproducible decision-to-contact traceability. The implementation is available at \href{https://anonymous.4open.science/r/Qubit-to-Motion-Quantum-Reinforcement-Learning-F7DC}{\textcolor{linkpink}{the anonymous repository}}.

Robot policies increasingly act through a small vocabulary of discrete latent actions learned from unlabeled video, yet training leaves open which motion receives which code. We find this choice highly unstable: two independently trained runs of Moto's latent motion tokenizer assign the same robot clips to nearly unrelated codes (adjusted Rand index 0.001), and evaluating one tokenizer in bf16 instead of fp32 reassigns 74% of its tokens. We use this freedom to impose the representational geometry of macaque ventral premotor cortex (PMv), recorded while a monkey drew shapes, on the tokenizer's code assignment, with a straight-through similarity loss trained alongside its reconstruction objective on robot video. In one training run per target, on five shape categories the loss never saw, the PMv-trained tokenizer reaches a representational similarity of 0.52 with PMv, against 0.20-0.26 for released and no-prior tokenizers and 0.42 for the 95th percentile of random kinematic encoders. Kinematic, language, and random targets do not transfer (-0.17 to -0.03). Two results limit the claim: a hand-picked kinematic descriptor, the stroke's initial direction, matches PMv better (0.65), and a label-shuffled PMv control scores 0.41. We specify the downstream manipulation test, on a CALVIN pipeline that reproduces Moto's published results, before running it.