VLA Gallery / All models / Humanoid whole-body controllers

Challenging-Terrain Transformer

Challenging-Terrain Transformer (Learning Humanoid Locomotion over Challenging Terrain)

UC Berkeley · 2024-10

VAHumanoid whole-body controllers1.4MRL policy

Pretrained via sequence modeling on trajectories from existing policies·model-based controllers·human motion capture, then fine-tuned with RL.

Paper

Learning Humanoid Locomotion over Challenging TerrainarXiv 2410.03654

Overview

Category

Humanoid whole-body controllers

Type

VA

Released

2024-10

Organization

UC Berkeley

Open weights

Unverified

Parameters

1.4M

Architecture

Backbone LLM

None

Vision encoder

—

Action head

Joint targets + PD gains (RL policy)

Chunk

Not disclosed

Control rate

50 Hz

Architecture (original summary)

1.4M Transformer, 16-step context; predicts joint target positions and PD gains at 50 Hz

Training

Pretraining accelerators

A100 × 4

Pretraining duration

Not disclosed

Pretraining steps & batch

300 epochs · global mini-batch 4,096

GPU/TPU-hours

Not disclosed

Fine-tuning

RL fine-tuning: A10 × 1, MuJoCo, 2,048 parallel envs × 24 steps per env × 2,000 iterations (~100M env steps), about 1 day

Data type

Simulation·synthetic / motion capture

Data scale

Pretraining trajectory count not disclosed; RL fine-tuning ~100M samples

Main datasets

Neural-network policy trajectories from prior work, model-based controller trajectories, existing human walking motion capture

Tasks & evaluation

Hiked over 4 miles of Berkeley trails, walked San Francisco streets with grades over 31%

Robot

Agility Digit

Inference

Inference device & speed

50 Hz

Inference memory

Not disclosed

Weights bf16 (calc.)

0 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Used (RL controller training) · Trajectory-sequence pretraining followed by MuJoCo RL fine-tuning (2,048 environments, about 100M steps)

Base control

Humanoid legged (RL policy drives joints directly) · Transformer takes velocity commands and directly outputs joint targets and PD gains at 50 Hz for rough-terrain locomotion (Digit)

Notes

Key features

Pretrained via sequence modeling on trajectories from existing policies·model-based controllers·human motion capture, then fine-tuned with RL. A rare case of a transformer directly handling a joint-level policy.

Notes

Pretraining wall-clock not disclosed. Fine-tuning compute recorded only in the fine-tuning column.

Verification

Confirmed