CHIRON — ANIMA Module

CRISP-Compliant ROS2 VLA Controllers — Production-grade robot control stack bridging Vision-Language-Action (VLA) models to physical manipulators via real-time compliant controllers.

Part of the ANIMA Perception Suite by Robot Flow Labs.

Overview

CHIRON is an orchestration layer for robot manipulation. It provides:

  • Cartesian impedance, operational space, joint impedance, and hybrid force-position controllers
  • VLA policy bridge for deploying learned policies to real hardware
  • Safety monitoring with velocity, torque, workspace, and singularity checks
  • Support for Franka FR3, KUKA IIWA14, and Kinova Gen3 manipulators

The policy network (ResidualMLP) maps robot state (8D) to action commands (8D), trained on the DROID 1.0.1 dataset (25.4M frames, 95K episodes).

Architecture

  • Model: ResidualMLP (input projection → 2 residual blocks → output projection)
  • State: 8D (7 joint positions + gripper)
  • Action: 8D (7 joint commands + gripper)
  • Parameters: 269,064
  • Hidden: [256, 256, 128] with LayerNorm + GELU + Dropout(0.1)

Exported Formats

Format File Size Use Case
PyTorch (.pth) pytorch/project_chiron_v2.pth 1.1MB Training, fine-tuning
SafeTensors pytorch/project_chiron_v2.safetensors 1.1MB Fast loading, safe
ONNX onnx/project_chiron_v2.onnx 1.1MB Cross-platform inference
TensorRT FP32 tensorrt/project_chiron_v2_fp32.trt 1.2MB Full precision (L4/A6000)
TensorRT FP16 tensorrt/project_chiron_v2_fp16.trt 1.2MB Edge deployment (Jetson)

Usage

import torch
from safetensors.torch import load_file

# Load model
state_dict = load_file("pytorch/project_chiron_v2.safetensors")
# Model: Linear(8, 256) → ResidualBlock×2 → Linear(256, 8)

# Or ONNX
import onnxruntime as ort
sess = ort.InferenceSession("onnx/project_chiron_v2.onnx")
action = sess.run(None, {"state": robot_state})[0]

Training

Parameter Value
Dataset DROID 1.0.1 (25.4M frames, 95K episodes)
Hardware NVIDIA L4 (23GB VRAM)
Precision bf16 mixed
Optimizer AdamW (lr=3e-4, wd=0.01)
Scheduler Cosine with 5% warmup
Batch size 2048
Epochs 30 (early stopped, patience=20)
Best val_loss 0.0041
Training time 44 minutes
Seed 42

See configs/train_droid.yaml for full config and logs/training_history.json for loss curves.

Simulation

CHIRON includes MuJoCo simulation backends for:

  • Franka FR3 (7-DOF)
  • KUKA IIWA14 (7-DOF)
  • Kinova Gen3 (7-DOF)

License

Apache 2.0 — Robot Flow Labs / AIFLOW LABS LIMITED

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading