Dex-SuperMimic — Reproducing SuperMimic on Dexterous Hand Manipulation

TL;DR

Dex-SuperMimic is a third-party reproduction of SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model SuperMimic adapted from humanoid whole-body control to in-hand dexterous manipulation on the Wuji Hand 1 and Wuji Hand 2 (plus a pen-spin task), built on mjlab + MuJoCo Warp. The repo ships three task configurations (hand1-reorient, hand2-reorient, hand2-pen-spin), a uv-managed install, Viser-based playback, and pretrained 8,000-iteration checkpoints distributed as GitHub Release attachments. Positioned downstream of SuperMimic’s three-phase recipe — collect real-sim rollouts → train differentiable world model → optimize policy against imagined rollouts, with periodic buffer refresh — but task-aware: cube reorientation uses absolute EMA control + contact/cage rewards + a translucent world-model ghost overlay in the viewer; pen spin tracks a packaged motion reference with reference-plus-residual control.

Key claims

Reproduces SuperMimic’s three-phase training loop on dexterous-hand tasks: (1) short real-sim rollout collection into a replay buffer, (2) recurrent multi-step world model trained on mixed normal + failure transitions, (3) policy + critic optimization along imagined rollouts with periodic buffer refresh from the real simulator [README §Overview].
Three task configurations covering two hands and two task families: Wuji Hand 1 cube reorientation at 20 or 50 Hz, Wuji Hand 2 cube reorientation at 20 or 50 Hz, Wuji Hand 2 pen spin at 50 Hz only [README §Tasks].
Task-aware implementation rather than uniform wrapper: cube reorientation uses command/goal observations + absolute EMA controller + contact & cage rewards + a translucent world-model ghost in Viser; pen spin tracks a packaged motion reference, uses a reference-plus-residual controller, and exposes the standard viewer without the cube-specific ghost overlay [README §Overview].
Linux x86_64 + NVIDIA GPU required for training and playback (CUDA 12.8 tested, Blackwell / RTX 50-series supported via MuJoCo Warp); macOS arm64 can install and import but training rejects non-CUDA devices [README §Requirements].
Pretrained checkpoints distributed as GitHub Release attachments at the 8,000-iteration mark for all three tasks, downloadable via curl and playable through dex-supermimic-play --checkpoint <path> [README §Pretrained Checkpoint].
Built directly on top of three upstream projects: InstinctMJ (mjlab-based sim framework foundation), Wuji MJLab (Wuji hand cube-reorientation + pen-spin task implementations), and supermimic_vibed (the SuperMimic reproduction this adaptation is derived from) [README §Acknowledgments].

Method

Three-phase training loop lifted from SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model SuperMimic and specialized to a dexterous-hand substrate. Phase 1 collects short rollouts from a MuJoCo Warp rigid-body simulator (via the mjlab backend) into a replay buffer. Phase 2 trains a recurrent-style multi-step world model on mixed normal and failure transitions pulled from that buffer. Phase 3 optimizes the policy and critic along imagined rollouts under the world model, with the replay buffer periodically refreshed from fresh real-simulator rollouts to keep on-policy signal in the training distribution. The task implementation diverges from SuperMimic in two task-aware ways: for cube reorientation the observation includes command/goal targets, control is an absolute EMA scheme, rewards include contact + cage terms, and the Viser viewer renders a translucent world-model ghost of the hand/cube alongside the real-simulator trajectory; for pen spin a packaged motion reference drives a reference-plus-residual controller and the viewer runs without the ghost overlay. Training is launched from a single CLI (dex-supermimic-train --task <hand1-reorient|hand2-reorient|hand2-pen-spin>), with the pen-spin task requiring an additional upstream motion-catalog download step.

Results

No benchmark numbers reported in the README; artifacts shipped are three pretrained checkpoints at the 8,000-iteration mark (model_hand1_reorient_8000.pt, model_hand2_reorient_8000.pt, model_hand2_pen_spin_8000.pt) [README §Pretrained Checkpoint].
Three demo GIFs embedded in the README documenting qualitative behavior: Wuji Hand 1 cube reorientation, Wuji Hand 2 cube reorientation, Wuji Hand 2 pen spin [README §Demos].
No sim-to-real transfer claims — training and playback are simulator-only, with Linux x86_64 + NVIDIA GPU as the only validated runtime target [README §Requirements].

Why it’s interesting

This repo is a concrete datapoint on how the SuperMimic recipe reaches its downstream communities: an outside reimplementation built in days on top of two chained community efforts (supermimic_vibed from zita-ch as the generic SuperMimic reproduction, Wuji MJLab for the hand tasks, InstinctMJ for the simulation framework), rather than released by the original authors or a well-funded lab. That mirrors the dynamic Andrew Carr — SuperMimic open reimplementations diverge during Bones Seed; four engineers compare notes called out directly — multiple parallel SuperMimic open reimplementations whose results diverge during Bones Seed and whose authors are comparing notes on failure modes — and extends it from humanoid whole-body control onto a dexterous-hand substrate the original paper did not evaluate. It’s an argument that the differentiable-world-model + BPTT recipe is portable by small groups across task families (whole-body loco → in-hand cube reorientation → packaged motion-reference pen spin), with the task-awareness living in reward shaping and controller choice rather than in a different training loop.

For Human-to-Robot Retargeting, Dex-SuperMimic sidesteps the interface-location debate the way the original SuperMimic does — no human-video retargeting step inside the loop, the training substrate is sim transitions on the target embodiment. The retargeting question would live upstream in whatever pipeline produces reference pen-spin motions (packaged motion catalog, downloaded via a bundled tool). For RL Environment Platforms, the repo joins the pattern Microduck RL (Microduck RL — RL training environments for a ~800 g open-source bipedal robot), OCBench (OCBench — A Controllable Robotic Manipulation Benchmark for BC and RL), and Cyclo (Cyclo — ROBOTIS Open-Source Modular Physical AI Framework) have been assembling — hardware-vendor-aligned (here: Wuji Hand) MuJoCo-based training stack packaged with pretrained checkpoints and viewer tooling, with the specific property that the training objective is model-based (SuperMimic BPTT) rather than the PPO-dominant pattern elsewhere in the cluster.