gr00t

FlashRT is a high-performance realtime inference engine for small-batch, latency-sensitive AI workloads. The flagship integration is production VLA control for Pi0, Pi0.5, GROOT N1.6, and Pi0-FAST. Also support llm e.g, qwen3.6-27B

Code for "ACG: Action Coherence Guidance for Flow-based Vision-Language-Action Models" (ICRA 2026)

Inspired by Tony Stark's robotic assistant Dum-E, the mission of this project is to create an intelligent, voice & vision enabled AI agent with robotic arm(s) capable of real-time human interaction, physical operations, and orchestration of a wide range of tools and services.

Experimental: Robot Control for Strands Agents.

1-click robot foundation model (VLA, world action model, LLM agent) simulation and evaluation on AWS GPU instances with AWS CDK

LoRA fine-tuning wrapper for NVIDIA GR00T N1.7 on a single 24 GB GPU

Monitoring & profiling toolkit for NVIDIA Jetson running Isaac-GR00T N1.5 TensorRT inference - Prometheus metrics, Grafana dashboards, per-engine latency profiling

Physical AI on the robot's own compute: GR00T through TensorRT on Jetson AGX Thor, Isaac Sim evaluation of a 6-DoF arm, ROS 2 on Orin NX, and a correction ledger kept on record. An EmbodiedEdge Labs project.

Open, safe operating system for humanoid robots — AI proposes, code decides, the OS protects.

Sim2Real pick-up for the JetRover arm in Isaac Lab: a PPO expert policy with task-space Delta IK, trained on 15 objects with material-based domain randomization, serving as a data engine that exports LeRobot/GR00T datasets for VLA training.

Track B SONIC SIM curriculum research fork: bounded eval, variant checkpoint provenance, and SIM-M2 gates for GR00T-WholeBodyControl

Levarage Humanoid WBC for simulated boxing Matches

Fine-tune NVIDIA Isaac GR00T N1.7 (3B VLA) on a single 24 GB RTX 4090 without root: patches, install script, MuJoCo evaluation, open datasets, 1800 evaluated attempts as JSON

NVIDIA Isaac GR00T N1.6 - A Foundation Model for Generalist Robots.

VIBTARVE: VLM-initiated behavior-tree auditing of robot VLA execution. Common-evidence replay, multi-model evaluation, and human-reviewed RoboCasa GR1 tabletop recordings from GR00T N1.6.

Research design: agentic RL + curriculum learning for GR00T humanoid whole-body control, bridging NVIDIA TAO's Cosmos/LoRA/agentic post-training stack. Fork of tao-skills-bank + GR00T submodules.

Bimanual handover on a Unitree G1_29DoF: replacing Dex3 hands with Brainco Revo2, retargeting a GR00T N1.7 policy, and recovering the task with simulation fine-tuning (Isaac Lab)

An NVIDIA Isaac GR00T N1.7 policy driving a simulated Franka Panda in MuJoCo: a closed perception-to-action loop from a language instruction and two camera views. Zero-shot, simulation only, with measured success rate and failure modes documented.

Execution-time supervisor for a fixed VLA on a humanoid hand: proprioceptive contact detection, per-finger grasp correction and compliant-mode selection — no tactile sensors, no VLA retraining. Unitree G1 + BrainCo Revo2, GR00T-N1.6.