VLA Gallery / All models / Mobile manipulation

SLIM

Horizon Robotics · 2025-01

VLAMobile manipulationUnder 19MRL policy

A quadruped+arm policy that performs long-horizon search·approach·grasp·carry·drop tasks from a single wrist RGB camera and language instructions, trained only with MuJoCo sim RL (privileged teacher→visual student) and deployed zero-shot on the real robot.

Paper

SLIM: Sim-to-Real Legged Instructive Manipulation via Long-Horizon Visuomotor LearningarXiv 2501.09905

Overview

Category

Mobile manipulation

Type

VLA

Released

2025-01

Organization

Horizon Robotics

Open weights

No

Parameters

Under 19M (low-level 3M + student high-level 16M; teacher 10M not deployed)

Architecture

Backbone LLM

None

Vision encoder

FiLM U-Net (auxiliary segmentation · depth) + ResNet 5 blocks, 160×90 RGB 5 frames

Action head

MLP actor: arm joint increments·gripper·forward/angular velocity (RL policy)

Chunk

None (1 step)

Control rate

High level 10 Hz, low level 50 Hz

Architecture (original summary)

160×90 RGB 5 frames → FiLM U-Net (auxiliary segmentation · depth prediction) + ResNet 5 blocks + Transformer 5 layers (8 heads), language via byte tokens 2-layer Transformer → MLP actor (512×3): arm joint deltas · gripper target · forward/angular velocity commands 10 Hz → PPO low-level locomotion policy leg joint targets 50 Hz

Training

Pretraining accelerators

N/A

Pretraining duration

N/A

Pretraining steps & batch

N/A

GPU/TPU-hours

Not disclosed

Fine-tuning

Sim RL (MuJoCo) only: teacher (multitask SAC) 4 GPUs about 1 week (120 envs·4M iter), student (SAC+KL distillation) 4 GPUs about 2 weeks (60 envs·1.3M iter), low level (PPO+ROA) 1 GPU about 3 days (5,000 envs·4B env steps); V100-class GPUs

Data type

Simulation·synthetic

Data scale

Sim RL: teacher about 60M·student about 40M env steps, low level 4B env steps; no real data

Main datasets

None (MuJoCo sim environment)

Tasks & evaluation

Real-world Lobby standard layout overall task success 78.3±5.8% (3 seeds×20 trials) vs expert teleoperation 75.0±5.0%; average 43.8 s (teleoperation 65.5 s); 400 episodes evaluated indoors and outdoors

Robot

Unitree Go1 quadruped + WidowX-250S arm, wrist RealSense D435 (RGB only)

Inference

Inference device & speed

Asynchronous inference on laptop RTX 3070Ti (i9-12900H, 16GB RAM): high-level 10 Hz·low-level 50 Hz; Raspberry Pi on the robot runs the drivers

Inference memory

Not disclosed

Weights bf16 (calc.)

0 GB

Where it runs

Laptop RTX 3070Ti

Reinforcement learning & base

Reinforcement learning

Used (RL controller training) · High-level teacher (multi-task SAC)·visual student (SAC+KL distillation)·low-level locomotion (PPO+ROA) all trained with MuJoCo sim RL, no real data

RL evidence

https://arxiv.org/abs/2501.09905

Base control

Quadruped (RL policy drives joints directly) · High-level policy outputs forward·angular velocity commands + arm joint increments·gripper → PPO low-level locomotion policy controls leg joints

Notes

Key features

A quadruped+arm policy that performs long-horizon search·approach·grasp·carry·drop tasks from a single wrist RGB camera and language instructions, trained only with MuJoCo sim RL (privileged teacher→visual student) and deployed zero-shot on the real robot. With under 19M parameters it achieved a 78.3% overall real-world task success rate, comparable to expert teleoperation (75.0%) while about 1.5× faster.

Notes

Added outside the list (2025 quadruped mobile manipulation VLA); deployed with no real data or fine-tuning; language encoder is a byte-token transformer without a pretrained LM; training GPU 'V100 or similar'; code only announced as forthcoming.

Verification

Confirmed