We introduce the first learned articulated body simulator based on a selective state space model (SSM), called PhysMamba. PhysMamba predicts next-frame full-body state from position, rotation, and joint-action history, without velocity inputs. We compare four architectures under partial- and full-observation inputs and three training protocols. The from-scratch rollout training protocol gives Mamba2 strong short- and mid-horizon accuracy under partial observation (s10 = 43 mm, 2/50 diverged), while the two-stage teacher-based rollout protocol stabilizes GRU but fails for Mamba2. With CUDA graph compilation, Mamba2 reaches 0.107 ms per frame (9,334 FPS) on an H100 GPU, within 1.1× of GRU's un-compiled throughput, adding under 1% latency to a 30 Hz HMR pipeline and enabling integration as a differentiable physics module for video-based mesh recovery.
Figures & tables
Figure 1: PhysMamba overview. Per-frame inputs (pose, rotation, gravity, action) are concatenated into a 285d vector and fed sequentially into a 4-layer Mamba2 SSM. The recurrent hidden state accumulates temporal context; the output head predicts the 358d next-frame physical state.
Hyperparameter
Value
dmodel
512
dstate
64
head_dim
64
nlayers
4
expand
2
Parameters
∼ 7M
Table 1: PhysMamba architecture configuration.
Model
285d (partial-obs)
430d (full-state)
Mamba2
29.2
37.5
GRU
36.6
29.2
MLP
49.4
49.9
MLP-LARP
88.3
86.0
Table 2: Single-step prediction (teacher forcing, val MPJPE in mm). Bold = best per column.
Model
Training
s1
s10
s50
s150
Div.
Mamba2 285d
TF
23.0
692
808
862
50/50
GRU 285d
TF
27.9
514
584
594
49/50
MLP-LARP 430d
Rollout
60.3
110
246
482
47/50
Mamba2 285d
Rollout
43.6
43.1
91.7
184
2/50
GRU 285d
TF → Ro
45.6
56.4
114.7
147
2/50
Mamba2 285d
TF → Ro †
105
174
203
242
33/50
Table 3: Autoregressive rollout (150 steps, MPJPE in mm, 285d input unless noted). “Div.” counts clips where any step exceeds 500 mm (50 val clips). † Gradient instability, see § 4.2 .
Figure 2: Per-step MPJPE over 150-step autoregressive rollout (GT actions, 50 clips). All models use 285d partial-obs input except MLP-LARP (430d full-state). All TF-only models diverge within 10 steps ( > 500 mm, not shown). Rollout-trained Mamba2 and two-stage GRU TF → Ro both remain stable across 150 steps.
Group
Model
Mean
Median
Trim5
left_leg
Mamba2
0.0730
0.0193
0.0336
GRU
0.1433
0.0528
0.0872
MLP
21.75
0.7189
0.9315
right_leg
Mamba2
1.2874
0.0190
0.0344
GRU
0.2134
0.0908
0.1386
MLP
21.27
0.8251
1.1155
Table 4: Per-body-group mean / median / 5%-trimmed-mean perturbation error (v1 rollout-trained checkpoints, 285d input). Mamba2’s mean is 3–33 000 × above its median on arm-containing groups, driven by ≈0.2% hard trajectories; GRU’s mean/median ratio stays near 2–3 × across all groups due to the bounded forget-gate update.
Learning motor control for muscle-driven musculoskeletal models is hindered by the computational cost of biomechanically accurate simulation and the scarcity of validated, open full-body models. Here we present MuscleMimic, an open-source framework for scalable motion imitation learning with physiologically realistic, muscle-actuated humanoids. MuscleMimic provides two validated musculoskeletal embodiments - a fixed-root upper-body model (126 muscles) for bimanual manipulation and a full-body model (416 muscles) for locomotion - together with a retargeting pipeline that maps SMPL-format motion capture data onto musculoskeletal structures while preserving kinematic and dynamic consistency. Leveraging massively parallel GPU simulation, the framework achieves order-of-magnitude training speedups over prior CPU-based approaches while maintaining comprehensive collision handling, enabling a single generalist policy to be trained on hundreds of diverse motions within days. The resulting policy faithfully reproduces a broad repertoire of human movements under full muscular control and can be fine-tuned to novel motions within hours. Biomechanical validation against experimental walking and running data demonstrates strong agreement in joint kinematics (mean correlation r = 0.90), while muscle activation analysis reveals both the promise and fundamental challenges of achieving physiological fidelity through kinematic imitation alone. By lowering the computational and data barriers to musculoskeletal simulation, MuscleMimic enables systematic model validation across diverse dynamic movements and broader participation in neuromuscular control research. Code, models, checkpoints, and retargeted datasets are available at: https://github.com/amathislab/musclemimic
Chengkun Li, Cheryl Wang, Bianca Ziliotto +4
EPFL, Switzerland · McGill University, Canada · Feil & Oberfeld CRIR-Jewish Rehabilitation Hospital Research Center, Canada +1
Tracking large human-motion datasets has become the dominant route to general, expressive humanoid whole-body control, but scaling across many motions is expensive. Current pipelines use model-free reinforcement learning, either merging per-motion specialists through lossy distillation, or brute-forcing a single policy with cluster-scale compute. SuperMimic amortizes learning across all motions with one differentiable world model of their shared dynamics, fit from logged transitions of a standard rigid-body simulator. The world model and the tracking policy train together in one loop, and backpropagation-through-time over short imagined rollouts gives the policy analytic gradients. Fast imagined rollouts and analytic gradients make policy learning efficient. In 3 hours on a single GPU, SuperMimic surpasses the return of a state-of-the-art motion tracker baseline trained for 7 days on 128 GPUs. It reduces GPU-hours by more than 1000× and fails on only 25 of 129,514 benchmark motions, versus 179 failures for the baseline. After 9.8 hours on two GPUs, SuperMimic achieves zero failures. From 1 to 264 hours of motion, the world model's prediction error stays stable, and GPU-hours to a target return grow slowly. Our policy transfers to a physical Unitree G1 across diverse and acrobatic motions. SuperMimic removes compute as the bottleneck to scaling humanoid whole-body control, making it practical to measure motion-tracking scaling laws.
We present PhysBrain 1.5, a unified model for understanding physical environments, generating actions, and predicting future states. Motivated by the physical loop of observation, interaction, and environmental change, we bring these capabilities into a common learning framework. Starting from a general vision--language model, we encode language responses, end-effector motion, and dense visual targets as discrete sequences and jointly optimize them with autoregressive next-token prediction. Pre-training draws its embodied supervision entirely from human interaction videos, using task-centered episodes to pair semantic and spatial context with recovered motion and subsequent observations. We then adapt the model through supervised fine-tuning on a mixture of human demonstrations, robot trajectories, and simulated experience. Across 28 embodied understanding benchmarks, our 8B model achieves an average score of 72.5, setting a new open-source state of the art and performing on par with leading proprietary models such as GPT-6-Astra and Gemini 3.6 Flash. It achieves the best open-source results on 14 benchmarks while retaining general multimodal capabilities. Beyond these understanding evaluations, qualitative examples show the model's ability to produce end-effector trajectories and predict future scenes through spatially aligned RGB, depth, and robot-mask outputs.