Launched this week
SimTooReal
Train robots in any sim. Ship to real world. Two commands.
6 followers
Train robots in any sim. Ship to real world. Two commands.
6 followers
The end-to-end instrument panel for robot learning engineers. SimTooReal covers the full pipeline: train in Isaac Lab/MuJoCo/Gazebo, stream 28+ live signals per iteration, auto-classify 50+ failure patterns, close the sim-to-real gap with 5-pillar scoring, collect & label data, fine-tune VLA/foundation models, run safety-gated deployments, and monitor live fleets — all from one platform. Two commands, 30-second setup, zero training loop changes.
Hey Product Hunt! 👋
I'm Vardhan, and I've been building robotics training infrastructure for the past year.
The frustration that led to SimTooReal: I was running the same policy on IsaacLab AND MuJoCo to cross-validate results — and I had five different terminal windows, three different logging formats, and no single place to compare reward curves side by side. Every team I talked to had the same duct tape.
W&B is great for ML broadly, but robotics has specific needs: body telemetry (CoM height, joint torques, base velocity), episode failure clustering, sim-to-real eval, and multi-environment dispatch. None of that fits cleanly into a generic experiment tracker.
So I built the missing layer. Two commands, your existing training script unchanged, full observability across every major sim.
The thing I'm most proud of today is Robot Studio — you describe a robot task in plain English, Claude generates the training project files, and you click "Run on GPU." Zero infrastructure setup. It's the fastest path from "I want to train a bipedal walker" to actual training curves I've ever seen.
We ran brutal benchmarks on an A10G: Ant-v4 on MJX 4096 envs hit 2,684 reward. We traced a bug where per-env Python loops triggered 40,960 GPU→CPU scalar transfers per update (10+ min/iter). Fixed to 3 transfers. Now it's fast.
💰 Pricing (all in USD):
→ Free — $0/forever. Live reward curves, GPU metrics, failure intelligence, MuJoCo + LeRobot support, 20 saved runs.
→ Pro — $25–$120/month. Full Isaac Lab + MuJoCo + Gazebo support, team dashboards, API keys, advanced validation gates, unlimited runs.
→ Pilot + Integration — $600–$2,400/month. 1-month enterprise pilot, custom pipeline build, Isaac Lab / MuJoCo integration support, failure validators, rollout gates, and team enablement.
I'd love your feedback on:
→ What sims/frameworks are you stuck on that we don't support yet?
→ What's the #1 thing missing from your current RL observability stack?
Try it free: simtooreal.com
CLI: pip install simtooreal-agent
Update from the maker — a few things I want to address directly based on questions we've been getting:
🔧 What exactly ships today (v0.42.1):
→ Python agent: pip install simtooreal-agent — wraps any training script, zero code changes, metrics stream in under 30 seconds
→ 28+ live signals per iteration: reward, entropy, KL, gradient norm, GPU temp/VRAM, FPS, ETA, curriculum stage, joint torques, and more
→ Failure intelligence: 50+ auto-classified error patterns (CUDA OOM, entropy collapse, reward hacking, joint limit violations, physics instability) with plain-English fixes
→ 5-pillar sim-to-real transfer scoring: joint distribution overlap, action coverage, physics alignment, behavioral consistency, sim fidelity — all must pass before hardware deployment
→ Free crash diagnosis at simtooreal.com/diagnose — paste any training log, get a root-cause analysis, no sign-in needed
→ Robot Hub: 40+ manufacturer-verified presets, URDF/MJCF/USD conversion across Isaac Lab, MuJoCo, Gazebo
→ VLA support: SmolVLA and Octo fine-tuning with LoRA adapters, ONNX export for edge deployment
→ LeRobot & MCAP/ROS 2 bag import for imitation learning pipelines
→ Fleet management: canary rollouts, per-unit live telemetry, one-click rollback
💰 Pricing (USD only, no ambiguity):
→ Free — $0/forever. 20 runs, core monitoring, MuJoCo + LeRobot, failure intelligence
→ Pro — $25–$120/month. Full Isaac Lab + Gazebo, team dashboards, API keys, unlimited runs, advanced safety gates
→ Pilot + Integration — $600–$2,400/month. 1-month enterprise pilot, custom pipeline integration, hands-on team enablement
🤔 The question we get most: "How is this different from W&B?"
W&B tracks scalars. We track robot state. The difference is joint distributions, sim-to-real gap measurement, episode failure clustering, physics safety gating, and fleet rollout control. You can't add a robotics plugin to a generic ML tracker and get this — you have to build from the robot up.
📬 If you're running Isaac Lab, MuJoCo, or Gazebo and want a 15-min walkthrough mapped to your specific pipeline: simtooreal.com/book-demo
Happy to answer any technical questions below.