RING (The One RING)

Allen Institute for AI·University of Washington · 2024-12

VLANavigationNot discloseddiscrete tokens

Builds a shape-agnostic indoor object search policy by imitation learning on expert trajectories of 1 million embodiments with randomized camera position·field of view·collider size etc. in sim, then RL fine-tuning with DD-PPO.

View architecture diagram in the gallery Paper ↗

Paper

The One RING: a Robotic Indoor Navigation GeneralistarXiv 2412.14401

Overview

Category: Navigation
Type: VLA
Released: 2024-12
Organization: Allen Institute for AI·University of Washington
Open weights: Yes
Parameters: Not disclosed

Architecture

Backbone LLM: None
Vision encoder: SigLIP-ViT-B/16 (frozen)
Action head: Linear actor-critic head (discrete action logits + value) (discrete tokens)
Chunk: None (1-step discrete action)
Control rate: Not disclosed
Architecture (original summary): Frozen SigLIP-ViT-B/16 image·text encoders (768→512 projection) → Transformer state encoder 3 layers + causal Transformer decoder 3 layers (512-dim·8 heads, 100-step context) → linear actor-critic head: shared discrete actions (MoveBase ±20cm, RotateBase ±6°·±30°, Done)

Training

Pretraining accelerators: 8 H100 80GB (IL)
Pretraining duration: Not disclosed
Pretraining steps & batch: 80k iterations · batch 240 trajectories (224 in Table 10), context 100 steps
GPU/TPU-hours: Not disclosed
Fine-tuning: DD-PPO RL fine-tuning: 8 H100 × 4 machines, 40M steps (64 parallel envs · rollout 128), time not disclosed
Data type: Simulation·synthetic
Data scale: Expert trajectories from 1M random embodiments·50k ProcTHOR houses (introduction mentions 150k houses)
Main datasets: ProcTHOR procedurally generated houses + Objaverse objects (AI2-THOR); HF allenai/ring-data
Tasks & evaluation: Sim ObjectNav average success rate over 5 embodiments 72.1% (FLaRe 44.8%, PoliFormer 55.4%, SPOC-2.3M 35.5%; +16.7%p on CHORES-S); real world Stretch 83.3%·Stretch (factory config) 72.2%·LoCoBot 80.0%·Go1 80.0% (average 78.9%)
Robot: Real world: Stretch RE-1 (2 camera configurations)·LoCoBot wheeled, Unitree Go1 quadruped, RB-Y1 wheeled humanoid (qualitative evaluation); sim: Stretch·LoCoBot·Unitree A1 etc.

Inference

Inference device & speed: Not disclosed (in human evaluation, a server runs the policy)
Inference memory: Not disclosed
Weights bf16 (calc.): Not disclosed
Where it runs: Not disclosed

Reinforcement learning & base

Reinforcement learning: Used (VLA RL post-training) · After imitation learning on expert trajectories, the policy itself is RL fine-tuned with DD-PPO (40M steps)
RL evidence: https://arxiv.org/html/2412.14401v2
Base control: Wheeled·Quadruped (Discrete motion actions) · Outputs discrete actions shared by all embodiments (MoveBase ±20cm·RotateBase ±6°/±30°·Done), executed by each robot's low-level controller

Notes

Key features: Builds a shape-agnostic indoor object search policy by imitation learning on expert trajectories of 1 million embodiments with randomized camera position·field of view·collider size etc. in sim, then RL fine-tuning with DD-PPO. Recorded 72.1% average success rate over 5 sim embodiments and 78.9% zero-shot on 4 real platforms including Stretch RE-1·LoCoBot·Go1.
Notes: Added outside the list: a representative 2024–2026 cross-embodiment·language-goal navigation generalist policy. Mismatches: IL batch 240 in main text vs 224 in Table 10, RL lr 2e-5 in main text vs 2e-4 in Table 10. Parameters·training time·inference device Not disclosed. HF checkpoint 587MB (ring_model_step_40356421).
Verification: Confirmed