ABot-C0: A General‑Purpose Control Intelligence Platform for Quadruped Robots

Why Quadruped Robots Need a Behavior Foundation

Human‑scale robots have advanced quickly using large‑scale motion‑capture datasets, but quadruped robots lack abundant data and suffer from cross‑modal distortion. ABot‑C0 addresses this gap by constructing a general‑purpose control intelligence base that unifies scalable motion data, universal tracking, robust terrain control, scene interaction, and a unified deployment pipeline.

Data Engine: Scaling Quadruped Motion Data

ABot‑C0 creates a multi‑source data pyramid comprising:

Motion capture – natural, biomimetic gait data.

Tele‑operation – high‑quality cold‑start demonstrations.

Manual design – high‑dynamic, expressive motions.

Video generation – text‑ and image‑prompted motion expansion for hard‑to‑capture behaviors.

The video branch first enforces robot identity consistency on the first frame, reconstructs 3‑D joint and root motion from monocular video, and then passes trajectories through semantic, geometric, and physical feasibility filters; only fully simulatable trajectories enter the training pool.

In total, the system yields 16,074 physically‑validated trajectories with language annotations , of which 7,488 originate from video generation and can be continuously expanded with compute.

Universal Motion Tracking: Data‑Driven Generalization

Training a single multi‑action RL policy on thousands of diverse dynamics leads to gradient interference. ABot‑C0 adopts an "expert‑to‑generalist" pipeline: train expert policies per motion, distill them into a Flow‑Matching policy via DAgger, and add a lightweight residual RL layer for robustness.

Two key designs improve data quality:

Dynamic‑aware motion selection – selects high‑value references based on physical feasibility, closed‑loop executability, and Flow confidence.

Manifold‑Calibrated Reference Condition (MCRC) – injects dynamic encodings from the motion manifold as structural cues for the policy.

Scaling experiments show a clear data scaling law: expanding training motions from 30 to 7,076 reduces unseen‑action MPJPE from 24.61 mm to 14.79 mm and raises success rate from 84.30 % to 88.54 %; the error gap between seen and unseen actions narrows from 10.17 mm to 2.41 mm. Adding MCRC further drops MPJPE to 12.53 mm and achieves a zero‑shot success rate of 91.02 %.

From Robust Gait to Full‑Terrain Mobility

ABot‑C0 builds locomotion in three layers:

Robustness & safety – implicit history representation + explicit state estimation; NP3O constrained RL eliminates torque, speed, and fall violations in 3.0 m/s sprint tests.

Biomimicry & omnidirectionality – Diff‑CAST diffusion‑based action priors preserve natural gait while symmetric commands enable forward, lateral, backward, and high‑speed motions.

Perception & terrain adaptation – privileged height‑map teacher policies are distilled into student policies via clean LiDAR memory; subsequent PPO fine‑tuning under noisy LiDAR and domain randomization lets the student rely on real sensor data.

On simulated terrain levels 0‑9, the full system attains an average success rate of 83.2 % and a highest‑level score of 7.8 , reducing dangerous foot‑slip from 31 % to 14 % compared to a perception‑only baseline.

Unified Deployment: Real‑Time System Integration

Real robots cannot reload policies per task, so ABot‑C0 runs motion control and tracking runners continuously in the background. A unified state machine arbitrates strategy mixing, safety checks, and recovery, using a single Robot I/O abstraction for both simulator and hardware. On the Tutu platform (camera + LiDAR), motor command loops run at 200 Hz while decision and policy inference run at 50 Hz .

Downstream Applications

With the unified stack, ABot‑C0 demonstrates two application domains:

Interactive companionship – the robot responds to open‑ended commands like “come over and shake hands,” combines voice and motion feedback, and can be interrupted safely.

Full‑terrain autonomous navigation – a high‑level planner generates waypoints, a local avoidance module converts them to velocity commands, and perception‑driven motion policies adjust gait using 3‑D point clouds to traverse slopes, unstructured ground, escalators, and irregular steps.

In a hand‑shake task evaluated on 2,048 IK reference trajectories, the average end‑effector error to the target is 1.18 cm , and the end‑to‑end pipeline links “walk‑to” and “perform action” into a continuous, repeatable behavior.

Future Outlook

The authors aim to extend ABot‑C0 toward a closed‑loop learning system where robots autonomously accumulate experience, identify failures, expand the motion library, and continuously improve policies, moving from isolated skills to self‑evolving embodied intelligence.

Technical Report: Behavior Foundations for Quadruped Robots: ABot‑C0 Technical Report

Paper: https://arxiv.org/pdf/2607.07370