RIO-2 is a two-rate WAM(World Action Model) built for robotics. RIO-2 is composed with a low-frequency visual-language S2 backbone and a high-frequency JEPA-diffusion S1 action policy. The model is designed to separate slow scene understanding from fast robot control:

• S2 refreshes visual-language context at low frequency.

• Bridge/compressor modules convert S2 context into compact action-conditioning tokens.

• S1 runs high-frequency action generation from cached S2 tokens and robot state.

• JEPA latent prediction provides an auxiliary future-action representation.

• A 10-expert S1 MoE residual path expands action capacity while keeping top-1 expert activation efficient.

RIO-2 uses JEPA-diffusion S1 action policy for general and flexible robot control in high frequency. S1 is MoE policy with 10 experts. Each expert is 100M parameter size. RIO-2's task memory maintains a small EMA latent memory over recent S2 context for longer-horizon task continuity.

S2 policy is inspires by allenai/MolmoAct2.

This repo uses Hub custom code. Pass trust_remote_code=True until RIO-2 is merged into Transformers.

RIO-2 is trained with allenai's opened datasets.

Key Configuration

state_dim: 6
action_dim: 6
action_horizon: 30
s2_token_count: 16
s2_width: 1024
s1_width: 384
s1_layers: 6
s1_heads: 8
s1_policy_mode: jepa_diffusion
s1_moe_num_experts: 10
s1_moe_top_k: 1
dtype: bfloat16

How To Load RIO-2 In Python

python

import torch
from transformers import AutoModel, AutoProcessor

model = AutoModel.from_pretrained(
"hoguai/RIO-2",
trust_remote_code=True,
torch_dtype=torch.bfloat16,
device_map="auto",
)

processor = AutoProcessor.from_pretrained(
"hoguai/RIO-2",
trust_remote_code=True,
)

model.load_s2_base(device="cuda")
model.refresh_s2(image, "pick up the red cube", force=True)
actions = model.act_fast(state, steps=2)

Runtime Pattern

RIO-2 is intended to run as a two-rate policy:

Refresh S2 when the scene or instruction changes, or at a low fixed rate.
Reuse cached S2 tokens inside the high-frequency control loop.
Call act_fast() repeatedly with the latest robot state.
Execute only the safe portion of the returned action chunk through an external safety controller.

Safety

RIO-2 outputs continuous robot actions and must not be connected directly to real hardware. Always place the policy behind a robot safety layer with joint limits, velocity/acceleration/jerk limits, workspace constraints, watchdog, E-stop, and a fallback controller.