We introduce Ψ0 (Psi-Zero), an open foundation model to address challenging humanoid loco-manipulation tasks. While existing approaches often attempt to address this fundamental problem by co-training on large and diverse human and humanoid data, we argue that this strategy is suboptimal due to the fundamental kinematic and motion disparities between humans and humanoid robots. Therefore, data efficiency and model performance remain unsatisfactory despite the considerable data volume. To address this challenge, \ours;decouples the learning process to maximize the utility of heterogeneous data sources. Specifically, we propose a staged training paradigm with different learning objectives: First, we autoregressively pre-train a VLM backbone on large-scale egocentric human videos to acquire generalizable visual-action representations. Then, we post-train a flow-based action expert on high-quality humanoid robot data to learn precise robot joint control. Our research further identifies a critical yet often overlooked data recipe: in contrast to approaches that scale with noisy Internet clips or heterogeneous cross-embodiment robot datasets, we demonstrate that pre-training on high-quality egocentric human manipulation data followed by post-training on domain-specific real-world humanoid trajectories yields superior performance. Extensive real-world experiments demonstrate that \ours\ achieves the best performance using only about 800 hours of human video data and 30 hours of real-world robot data, outperforming baselines pre-trained on more than 10× as much data by over 40% in overall success rate across multiple tasks. We will open-source the entire ecosystem to the community, including a data processing and training pipeline, a humanoid foundation model, and a real-time action inference engine.
Figures & tables
Fig. 2: Model Training and Deployment : First, we pre-train the VLM on the EgoDex [ 20 ] dataset to autoregressively predict the next-action tokens in the task space. Then, we post-train the flow-based action expert using robotic data to predict action chunks in the joint space. Finally, we implement a real-time chunking mechanism that leverages the lower-body controller to achieve smooth whole-body control.
Fig. 3: MM-DiT for VLA: Comparison of MM-DiT architecture with naive DiT. τ is the flow timestep and VL and A denotes hidden states of the vision-language and action respectively.
Fig. 4: Real-Time Chunking: Given that the previous action is being executed (yellow line), the next action can diverge significantly (cyan line) without RTC, which leads to control jitter. With RTC (red line), the divergence between two consecutive actions is strongly suppressed, resulting in smoother and more stable behavior.
Fig. 5: Real-Robot Teleoperation Setup: We use MANUS gloves for dexterous hand retargeting; a VR headset and wrist trackers capture upper-body poses for inverse kinematics, while waist and foot trackers provide high-level locomotion commands.
Fig. 6: Real-World Task Setup: We evaluate Ψ0 on eight diverse long-horizon dexterous loco-manipulation tasks involving manipulation, whole-body motion, and locomotion. The task instruction is overlayed on the task images and each sub-task is denoted with marker for better visualization. Our policy rollout videos are included in the Supplementary Materials.
Fig. 7: Real-World Benchmark: Evaluation results of policies across our eight tasks, showing task-wise success rates (%) (left) and aggregated skill-level success rates (%) (right). The task descriptions are shown in Fig. 6 . Detailed results for each task including all sub-task progress are included in the Supplementary Materials .
Pre-Training
Post-Training (On HE)
Real-Time Chunking
MM-DiT Action Head
Naive DiT Action Head
Right-Arm Pick-n-Place
Left-Arm Pick-n-Place
Dual-Arm Carry
Overall Success Rate
EgoDex
HE
✗
✗
✗
✗
✗
✓
1/10
1/10
1/10
0/10
✗
✗
✗
✗
✓
✗
9/10
2/10
3/10
2/10
✓
✗
✗
✗
✓
✗
8/10
6/10
6/10
6/10
✓
✓
✗
✗
✓
✗
8/10
8/10
9/10
8/10
✓
✓
✓
✗
✓
✗
9/10
9/10
10/10
9/10
TABLE I: Ablation Studies . We study the effects of pre-training, post-training, and real-time chunking on a dual-arm long-horizon task which consists of three steps: right-arm pick and place, left-arm pick-and-place and dual-arm lift.
Reconstruction L1 Loss
Avg Token Length
Before
5.83×1e−4
2.08
After
1.95×1e−4
13.04
TABLE II: Fast Tokenizer. Comparison of reconstruction loss and average token length before and after training. Boldface indicates the best performance.
Fig. 8: Training-Time RTC. Diffusion timesteps and loss calculation in training.
Fig. 9: Real-Time Action Chunking System Design. The system consists of a client (observation collection and action execution) and a server (control and inference). The Control Loop (30Hz) coordinates observation updates and action dispatch, while the Inference Loop runs asynchronously to compute the next action chunk when t≥smin , enabling seamless chunk transitions without inference-induced interruptions.
Fig. 10: Single Operator Teleoperation Framework. Our framework maps human upper-body motions to robot arm and hand control via retargeting and multi-target IK, while lower-body pose is generated through an RL-based policy.
Task 3 Task 2
Descriptions
Pick the bottle, turn around, and pour into cup
Spray the bowl with water, wipe clean, and fold it up
Grasp
Move
Pour
Place
Overall
Grasp
Pull
Spray
Put
Overall
Diffusion Policy
0/10
0/10
0/10
0/10
0/10
0/10
0/10
0/10
0/10
0/10
ACT
0/10
0/10
0/10
0/10
0/10
3/10
2/10
4/10
3/10
1/10
InternVLA-M1
0/10
0/10
0/10
0/10
0/10
0/10
0/10
2/10
4/10
0/10
EgoVLA
4/10
6/10
1/10
2/10
1/10
0/10
0/10
0/10
0/10
0/10
TABLE III: Real-World Benchmarking: We provide a detailed report of real-world benchmarking results, including sub-task progress. Each task consists of three to five subtasks, and a trial is counted as successful only if all subtasks are completed. Boldface indicates the best performance, while underlining denotes the second-best performance.
GR00T-N1.6
Pick the dumpling
Pick the hippo
Carry the box
Overall SR
w/o RTC
10/10
7/10
9/10
7/10
w/ RTC
6/10
7/10
10/10
6/10
TABLE IV: GR00T with RTC. We study the effect of RTC on the GR00T baseline. The task consists of three steps. It achieves comparable performance on GR00T with and without RTC.
Experiment 1
Pick the dumpling
Pick the hippo
Carry the box
Overall SR
Baseline ( Ψ0 )
9/10
9/10
10/10
8/10
Variant (10% EgoDex)
6/10
1/10
5/10
1/10
Experiment 2
Grasp bottle
Wipe the bowl
Stack up
Overall SR
Baseline ( Ψ0 )
10/10
9/10
7/10
7/10
Variant (10% EgoDex)
9/10
10/10
7/10
6/10
TABLE V: Ablation of Pre-Training on 10% EgoDex. We found that using 10% of EgoDex perform worse than the baseline Ψ0 , demonstrating the efficacy of full EgoDex pre-training.
Experiment 1
Pick the dumpling
Pick the hippo
Carry the box
Overall SR
Baseline ( Ψ0 )
9/10
9/10
10/10
8/10
Variant (HE)
9/10
4/10
10/10
4/10
Experiment 2
Grasp bottle
Wipe the bowl
Stack up
Overall SR
Baseline ( Ψ0 )
10/10
9/10
7/10
7/10
Variant (HE)
10/10
9/10
4/10
4/10
TABLE VI: Ablation of Pre-Training on HE. We discover that the HE variant achieves high performance on tasks that do not require fine-grained manipulation; however, it still lags behind our baseline on subtasks requiring more precise manipulation.
Fig. 11: Multi-Task Fine-Tuning. Joint fine-tuning (Cyan) across multiple tasks leads to a performance drop on individual tasks (Pink).