Humanoid

The final guide in the series: current cuRobo issues, Jetson deployment risks, and safety checklists for robot arms and Unitree G1.

VIMA series capstone: adapt from tabletop 6-DoF to Unitree G1 23-DoF — freeze backbone, swap action head, bridge data gap with teacher-student and LoRA fine-tuning.

PlotJuggler Custom Function Editor writes Lua computing per-joint power (τ × dq) and tracking error — detecting overloaded joints /lowstate doesn't expose.

A ZMQ bridge from MuJoCo/Python to PlotJuggler to stream G1 sim state live, sync VideoViewer with hardware, and compare sim-to-real gaps per joint/gait phase.

A deployment-oriented guide to Unitree G1 whole-body workflows with cuRobo extra_links, tool frames, MotionRetargeter, IK/MPC, and BVH/SOMA data.

Install ros-humble-plotjuggler-ros, configure CycloneDDS for Unitree G1, and stream /lowstate and /sportmodestate live in under 15 minutes.

Load G1 MCAP recordings with the DataLoadMCAP plugin, drag 23 joint channels into a synced multi-panel grid, save layout as XML, and plot phase portraits.

ToolboxQuaternion converts G1 IMU quaternions to roll/pitch/yaw, ToolboxFFT detects gyroscope resonance, and Moving Average smooths foot_force data.

SOMA-X v0.2 unifies body models, pose conversion, LODs and pose inversion for physical AI, humanoid simulation and robotics.

Connect SONIC to MotionBricks with VQVAE, pose/root models, synthetic training, and ZMQ v4 token_state.

A beginner-friendly guide to MPC-RL: MPC-shaped PPO rewards, repo setup, training, inference, and results for humanoid locomotion and pushing.

Follow the OpenWBT pipeline: VR wrist pose, G1_29_ArmIK, action_hl_idx clipping, and G1 upper body control in MuJoCo.

A practical G1 sim2real checklist covering DDS, ROS_DOMAIN_ID, OpenWBT, .pkl logs, PlotJuggler, damping and torque guards.

Using cuRobo for Unitree G1 arm-only workflows with locked legs/waist, joint maps, wrist tool frames, self-collision, and command gates.

Compare robot dataset licenses, commercial training risk, closed fleet data moats, and the likely open vs closed landscape in 2027.

A practical playbook for small robotics teams: SO-101 data, LeRobotDataset v2, pretrained VLA fine-tuning, and HF datasets.

How to turn a few human demos into thousands of robot trajectories with MimicGen, GR00T-Mimic, RoboCasa, and sim-real co-training.

Robot data does not scale by raw demo count alone. This guide explains diversity power laws and when more data becomes wasted spend.

Compare AgiBot, Unitree XR, ALOHA 2, Mobile ALOHA, and GELLO to choose a real-world teleoperation data stack by budget.

How Ego4D, EPIC-Kitchens, web video, LAPA, GR00T N1, EgoMimic, and EgoHumanoid turn human video into robot pre-training data.

Compare Phantom and pi0.5 to decide when human videos become training data, derived robot data, or benchmark evidence.

Close the series with the VLA lifecycle from LeRobot datasets and norm stats to pi0.5 checkpoints, policy serving, and robot actions.

A beginner-friendly map of humanoid data: raw files, LeRobot, Isaac Lab, Humanoid Everyday, and who controls value at each layer.

A practical EgoHumanoid walkthrough: PICO/ZED capture, raw HDF5/SVO2, five processing stages, and consent/licensing audit points.

How EgoHumanoid uses depth warping and inpainting to turn egocentric human video into robot-compatible observations.

Final series post: start the FastAPI inference server, connect G1 via SSH tunnel, send action commands, run arm VLA and locomotion simultaneously — with safety checklist and debug guide for common real-hardware failures.

Train SONIC task-general trackers, map paper metrics to code paths, and export successful GRAIL rollouts.

Dissect GRAIL's two asset paths: procedural terrain and Hunyuan3D text-to-mesh for loco-manipulation.

Run gen_2dhoi for cordless_drill: Bullet settling, scale/prompt refinement, scene rendering, and Kling videos.

A beginner guide to recon_4dhoi: human pose, masks/depth, FoundationPose, HOIOptimizer, filtering, and visualization.

Trace GRAIL retargeting: SMPL-X to G1, hand actions, table geometry, lift thresholds, and BPS embeddings.

Fine-tune unifolm-vla from the public Qwen2.5-VL-7B-Instruct checkpoint (since Unifolm-VLM-0 is not public) — covering the official 8-GPU DeepSpeed approach and a single-GPU QLoRA workaround for beginners.

Architecture and design principles for extending UMI from upper-body manipulation to whole-body humanoid data collection by combining it with mocap or VR teleop — practical analysis from HumanPlus, OmniH2O, and Open-TeleVision.

Closing the loop: domain randomization in Isaac Lab, actuator modeling for sim2real gap, deploy to real G1, benchmark with humanoid-bench, plus a debug guide.

Deep dive into VLA and Whole-Body Control GitHub repos from China and the open community — Unitree (unifolm-vla, xr_teleoperate, unitree_rl_gym), THU RDT-1B, RobotEra humanoid-gym, and benchmark tools.

Deploy the GR00T-WBC stack (GEAR upper body 50Hz + SONIC locomotion 200Hz) on G1 with a fine-tuned GR00T N1 checkpoint — joint mapping, PD tuning, safety.

Deep dive into VLA and Whole-Body Control GitHub repos from US companies and labs — NVIDIA Isaac-GR00T, GR00T-WholeBodyControl, Physical Intelligence openpi, Berkeley HumanPlus, Stanford TeleVision, OpenHelix, and EgoHumanoid.

A practical guide to humanoid robot cameras: head RGB-D, wrist cameras, RealSense, OAK-D, latency, ROS 2, calibration, and affiliate placement.

3-step guide to converting teleoperation data: from JSON (xr_teleoperate) to LeRobot V2.1, to HDF5, to RLDS format ready for unifolm-vla training — with dataset verification at each step.

A practical roadmap for humanoid robot hardware: compute, actuators, cameras, IMU, power, safety, logging, and natural affiliate placements.

Fine-tune GR00T N1 (2B params) with a LeRobot dataset from G1 — GPU requirements, training config, avoiding overfitting, evaluating checkpoints before deploy.

A complete map of active GitHub repositories from humanoid companies and research labs working on VLA (Vision-Language-Action) and WBC (Whole-Body Control) — from NVIDIA to Unitree.

A practical comparison of Jetson Orin Nano, Orin NX, and cloud GPUs for humanoid robots across ROS 2, cameras, VLA inference, logging, training, and budget.

Collecting data for GR00T N1 via two tracks: Isaac Lab simulation teleop (no robot needed) and real xr_teleoperate on G1 — both producing LeRobot format.

Step-by-step guide to setting up xr_teleoperate with Meta Quest 3, connecting G1, recording teleoperation demos, and collecting high-quality data for VLA training on real hardware.

How GR00T N1 (VLA at 6Hz) combines with GR00T-WBC (200Hz) to control Unitree G1 — decoupled architecture, why end-to-end fails, how to adapt for other robots.

Overview of the 3-repo architecture — unifolm-vla (VLA), xr_teleoperate (data collection), unitree_rl_gym (locomotion) — and how they connect into a whole-body control pipeline for Unitree G1 on real hardware.

Detailed guide to SUGAR (arXiv 2026.05) — humanoid loco-manipulation framework that learns from real human videos, fully open-source, beginner-friendly end-to-end.

Detailed guide to Booster Gym — open-source end-to-end RL framework training Booster T1 humanoid walking from teleop to real-world deployment.

Reproduce LIFT (ICLR 2026): pretrain SAC in MuJoCo Playground, then sim2sim fine-tune a physics-informed world model in Brax for Booster T1 and Unitree G1.

Reproduce SteadyTray on IsaacLab — train a Unitree G1 humanoid to carry a tray without dropping objects using a two-stage Residual RL architecture.

Complete pipeline for deploying RL locomotion policies to real humanoid robots — domain randomization, system ID, safety, and Unitree SDK.

Parkour for humanoid robots — jumping obstacles, climbing stairs, navigating rough terrain with RL and terrain perception.

Combine locomotion and manipulation: carry boxes, push carts, open doors while walking, with multi-objective reward design.

Explore H1-2 with dexterous hands and improved actuators, how hardware changes affect locomotion, and loco-manipulation basics.

Extend H1 policy to running at 2+ m/s, sharp turning, lateral walking, and a multi-gait command-conditioned policy.

Train locomotion for the full-size Unitree H1, compare with G1, and adjust rewards and PD gains for a 1.8m humanoid.

Detailed analysis of 3 most popular humanoid robot platforms — specs, pricing, SDK, and practical applications.

Extend G1 walking policy to complex terrain with curriculum learning, domain randomization, and teacher-student training.

Step-by-step tutorial to train a walking policy for Unitree G1 with PPO in Isaac Lab, from config to evaluation.

Detailed analysis of 12+ reward terms for humanoid walking, reward hacking avoidance, and reward curriculum from standing to walking.

Learn why humanoid locomotion is hard, MDP formulation for bipedal walking, and how to set up Isaac Lab environment for Unitree G1.

Humanoid robot market at $38B by 2035 — hiring trends, essential skills, and learning roadmap for Vietnamese robotics engineers.

Techniques for controlling humanoids that both move and manipulate objects — decoupled control, teleoperation, and imitation learning.

Guide to training locomotion policies for humanoid robots with Humanoid-Gym — from reward design to zero-shot sim-to-real transfer.

Whole-body Model Predictive Control with MuJoCo and iLQR — from contact dynamics theory to real-time deployment on humanoid robots.

From Inverse Kinematics to operational space control — theory and Python code for controlling humanoid robot balance and task execution.

Detailed comparison of Atlas, Optimus Gen 3, Figure 02, Unitree H1/G1/H2, Fourier GR-2, and 1X NEO -- specs, pricing, SDK and ROS 2 support.

Overview of humanoid robots — bipedal walking, whole-body control, and platforms like Tesla Optimus, Figure 02, and Unitree H1.