MimicLite is an efficient, general humanoid motion-tracking system that can train a deployable policy in 3 hours on 8 RTX 4090 GPUs while retaining competitive tracking quality. Under a matched MuJoCo evaluation, MimicLite improves global root tracking over SONIC while achieving comparable local tracking accuracy. The same policy supports low-latency Pico-driven teleoperation and highly dynamic motion tracking on a physical Unitree G1.
The technical report is available at mimic-lite.pdf.
This repository is the project landing page. Training, evaluation, dataset conversion, and deployment instructions are maintained in their respective repositories:
| Component | Repository | Contents |
|---|---|---|
| MimicLite | EGalahad/mimic-lite |
Training, evaluation, policy export, task configs, and learning code. |
| Training framework | Agent-3154/active-adaptation |
Simulation backends, distributed launchers, environments, and shared infrastructure. |
| Motion data toolkit | EGalahad/any4hdmi |
Motion conversion, validation, visualization, and dataset tooling. |
| Deployment runtime | EGalahad/sim2real |
ONNX inference, MuJoCo sim2sim, Pico teleoperation, and Unitree G1 deployment. |
The released checkpoint set contains MimicLite v1.1 and the original Huge and Base PPO releases. The wall-clock column reports training time on RTX 4090 GPUs.
| Policy | Actor hidden dimensions | Parallel environments | Checkpoint | Wall-clock time |
|---|---|---|---|---|
| MimicLite-v1.1 | [1024, 1024, 1024] |
16 × 16000 (15000 finetune) |
ece66d1a |
15 h 29 min |
| MimicLite-Huge | [1024, 1024, 1024] |
32 × 8192 |
xua2csee |
3 h 30 min |
| MimicLite-Base | [1024, 1024, 1024] |
8 × 8192 |
iij0q0b5 |
4 h 20 min |
Original training-time sources: Huge 55ie49o5 and Base iij0q0b5.
Download the MimicLite v1.1 deploy ONNX and YAML from the shared sim2real artifacts.
For a fair comparison, we report the motion-lookahead latency required by each policy, defined by its furthest required future-reference frame. All values use the shared 50 Hz reference-motion contract.
| Policy | MimicLite-v1.1 | MimicLite-Huge | MimicLite-Base | BFM-Zero | SONIC | SONIC-v1.1 | SONIC low-latency | HoloMotion | TeleopIT | Humanoid-GPT | HEFT | TWIST2 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Motion-lookahead latency | 0.08 s | 0.08 s | 0.08 s | 0.12 s | 0.90 s | 0.90 s | 0.18 s | 0.20 s | 0.00 s | 0.02 s | 0.12 s | 0.00 s |
Released training datasets are collected in the any4hdmi Hugging Face collection. The BONES-SEED dataset is the exception: to respect its license and redistribution terms, users obtain it from the original source, while EGalahad/any4hdmi provides only the conversion scripts and processing tools.
The sim2real runtime provides a modular observation interface that separates policy-specific input construction from the shared deployment runtime. Integrating a policy requires only an observation class and a YAML specification; the inference, simulator, and robot interfaces remain unchanged. This common path supports integrated MuJoCo evaluation and real-robot execution for MimicLite, HEFT, TeleopIT, Humanoid-GPT, BFM-Zero, SONIC, and TWIST2. Policy inference is decoupled from robot I/O through interchangeable MuJoCo and physical Unitree G1 backends.
This integration repository is released under GPL-3.0-or-later. Component repositories retain their own histories and license files; verify dataset and component licenses before redistribution.