World Action Model

From skills
to new tasks

Grasping, turning and pulling can serve many tasks. We study reusable atomic skills and how to organize them around a goal and the current state of the world.

Explore research Evaluation levels

Atomic skill composition · Research concept

Target stateThe door is open

Current relationThe hand grasps the handle

Reusable action skills

GraspTurnPullMoveRelease

Conceptual example: organize skills around a target state and update observations after each step.

HyperNet · Method foundations

Networks that
generate networks

A hypernetwork produces the parameters of another network. This offers a way to generate representations or control policies from a condition, and informs our research into reusable action skills.

HyPoGen brings the structure of optimization into policy generation. The new HyperNet manuscript builds on this principle, using attention modules to scale to implicit representations of images, 3D shapes and NeRFs.

Conditioning input

Generate network parameters

Target network

The manuscript tests 2D and 3D INR generation on five datasets. Skill composition is a research direction; robot experiments and generalization benchmarks test control and transfer separately.

Scaling Optimization-Oriented Hypernetworks for Implicit Neural Representations · Research manuscript

HyPoGen · ICLR 2025

DISC · Control policies

A different goal.
A generated policy.

DISC generates a task-specific policy from a language instruction. During execution, the policy reads visual observations and robot state to produce control actions.

Read the DISC paper Research code

DISC real-robot experiment showing the same apple placed in different containers according to the instruction — Different instructions specify different placement goals in the same scene. These images show DISC’s real-robot experiment.Ren et al., DISC, Fig. 3 · CC BY 4.0

≈30MTrainable parameters

With about 30 million trainable parameters, DISC achieves results close to π₀.₅ in the reported LIBERO experiments.

Success rates, parameter counts and experimental conditions

The counts use different scopes. DISC’s 30M counts trainable parameters, excluding the frozen T5-small language encoder; its ResNet-18 visual encoder is fine-tuned. π₀.₅’s approximately 3.3B counts total model parameters.

Spatial / Object / Goal results use joint training, with 50 evaluations per task. π₀.₅ is fine-tuned from its pretrained checkpoint.

Trainable parameter count Benchmark results π₀.₅ architecture π₀ · 3B + 300M

Success rate in the experiments reported by DISC
Benchmark	DISC	π₀.₅
LIBERO Spatial / Object / Goal (overall success rate)	98.7%	98.3%
LIBERO-90	94.3%	95.7%

View the DISC architecture

DISC architecture: a language encoder and hypernetwork generate the policy weights; an observation encoder supplies state inputs to the generated policy. — Ren et al., DISC, Fig. 2 · CC BY 4.0 · Image license

The Imitator Game · Generalization evaluation

Make generalization
testable

Xunzhe Zhou and collaborators introduce four levels of human-video imitation. Each increases the difference between a demonstration and the robot’s scene, testing whether the same goal can be achieved under new conditions.

Xunzhe Zhou et al. · CoRL 2026

The same goal can require different actions

We make the scope of generalization explicit: changes in position, appearance and function. The Imitator Game supplies graded tasks, paired demonstrations and evaluation protocols for comparisons under defined conditions.

Explore the official project Read the paper

Watch the full project overview2:51