SIMPLE: Simulation-Based Policy Learning and Evaluation for Humanoid Loco-manipulation
1. Paper overview
In one sentence: SIMPLE separates humanoid contact physics from photorealistic rendering to make demonstrations and policy comparisons reusable, at the cost of expensive rendering and still-limited evidence for broad physical transfer. e02e03e05e08e09e10e12e13
| At a glance | What to know |
|---|---|
| Research problem | Source description Humanoid manipulation couples balance, locomotion and hand–object contact. Physical evaluations are costly and difficult to reset consistently, while existing platforms often separate locomotion from object interaction. The authors seek repeatable policy comparison and scalable demonstration collection within one whole-body environment. e02 |
| Core mechanism | |
| A key reported result | Zero-shot real-robot pick-and-place and handover: Physical Pick & Place: 8/10; physical Handover: 8/10. Task successes in ten trials.. One Psi0 policy fine-tuned only on simulation data; physical evaluation under similar environmental settings. Simulation: 9/10 and 10/10, respectively. Shows downstream transfer without real-world fine-tuning on these two tasks. It does not imply no real-data pretraining or broad environment generalization. e12e15 |
| Reading caution |
Core contributions
Figure 2. Behavior collection, visual replay and policy evaluation are separate stages. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the columns from left to right. The scripted route begins with a grasp pose and cuRobo planning; the PICO route maps operator motion through multi-target inverse kinematics. Whole-body control supplies the execution support shown beneath them. In the middle column, arrows carry raw trajectories through Isaac Sim replay, filtering and a success-only selection. The icons enumerate variations in object pose, lighting, language, viewpoint, material and grasp pose. The right column groups evaluation changes into backgrounds/distractors, materials/lighting and spatial poses. This is a workflow for producing and evaluating behavior, rather than a diagram of one learned policy network. e03e05e06e17
What it supports. The useful separation is between collecting physically executed trajectories and producing varied visual observations from them. VR collection can use the lightweight MuJoCo renderer, while expensive Isaac Sim rendering happens later. The original diagram also makes the success-only selection explicit; rendered demonstration data therefore passes through a filtering stage.
Where the evidence stops. The figure and appendix use Levels 0–2; Section 3.5 uses 1–3 for the corresponding progression. The controller box lists AMO/SONIC/GROOT, whereas the controller appendix details AMO and SONIC variants. Neither difference is silently resolved here.
2. Motivation
2.1 The problem and the proposed response
Humanoid manipulation couples balance, locomotion and hand–object contact. Physical evaluations are costly and difficult to reset consistently, while existing platforms often separate locomotion from object interaction. The authors seek repeatable policy comparison and scalable demonstration collection within one whole-body environment. e02
2.2 What this reading follows
A humanoid benchmark must test more than whether an arm reaches an object: it must also preserve balance, navigation and contact while presenting useful visual observations. SIMPLE approaches this through a simulator pairing. MuJoCo executes the physical interaction; Isaac Sim supplies photorealistic images from synchronized states. Motion planning and VR operators generate demonstrations, which are rendered offline for policy adaptation. The five visuals below follow that pipeline into policy comparisons, collection efficiency, controlled data ablations and physical transfer. Read the final transfer table as evidence from two tasks under similar settings, while keeping the broader claim about real-world policy rankings separate. e02e03e05e08e09e10e12e13
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Benchmarks & simulators |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded benchmark/simulator category and Not applicable architecture, prediction paradigm and quadrant are supported. SIMPLE composes simulators, controllers and external policies; it does not propose a unified learned future/action predictor. Including DreamZero as a baseline does not assign DreamZero’s architecture to the testbed. e03e08e14e16
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
5. Method in detail
5.1 Build valid motion before multiplying visual conditions
Begin with the object rather than the neural policy. SIMPLE needs a mesh that can collide stably in MuJoCo and appear consistently in Isaac Sim. CoACD and USD conversion supply those two representations. Free-drop simulation identifies resting poses; BoDex synthesizes grasps for caching; CuRobo uses selected grasps during task decomposition. VR demonstrations provide another route when scripted planning struggles with dexterous behavior. The recorded motion is subsequently replayed for rendering, with success-only filtering shown in the pipeline. The architectural benefit, as a reader interpretation, is that collecting behavior and varying its appearance need not consume the same resources at the same time. But changing objects or spatial poses requires preserving the validity of the demonstration; the paper does not quantify how much candidate data the filtering discards. e03e04e05e06
5.2 Follow the action through the controller boundary
At the policy boundary, SIMPLE supplies RGB observations, instructions and a 32-dimensional state comprising hand, arm, waist and base-height coordinates. An action adds four navigation commands to the corresponding 32 targets. This representation lets different policies connect to the same environment, but their output still needs controller execution. AMO separates upper-body tracking from lower-body locomotion, while SONIC offers whole-body tracking variants. MuJoCo advances the physical interaction and Isaac Sim receives the resulting states for visual feedback. Remote policy clients buffer predicted action chunks for stepwise execution. The training setup changes between baselines: Psi0 uses chunks of 30, ACT uses 100, and GR00T runs sequential inference. A common interface therefore standardizes communication without making the learned policies or their execution schedules identical. e07e14e15e16e17
5.3 Separate robustness, data quantity and physical transfer
The most informative ablation is the small BendHandover comparison because its two rows keep the training-example count fixed. Replacing half the Level-0 data with Level-1 examples changes the hardest evaluation set from five to eight successes, but the middle set falls by one success. That is a specific generalization tradeoff. The XmoveBendPick rows address a different question: whether more demonstrated behavior helps at a fixed fine-tuning-step count. Their improvement cannot be attributed solely to visual diversity. Finally, Table 5 changes the evaluation domain to a physical robot while keeping downstream training in simulation. As a reader interpretation, these experiments support three distinct uses of the testbed; none alone validates the broader statement that it predicts real-world rankings for many policies. e08e10e12
5.4 Training and inference
During training
The benchmark uses single-task fine-tuning. Psi0 is adapted for 40,000 steps on eight A100 GPUs: batch 16 per device, cosine learning rate 0.0001, 1,000 warmup steps, bf16 and action chunks of 30. Its action expert uses flow matching; SIMPLE introduces no common new learning objective. e08e15
DreamZero receives rank-4 LoRA adaptation for 40,000 steps on eight A100s. ACT uses chunks of 100. GR00T uses 20,000 steps and sequential inference; InternVLA freezes its VLM. Different checkpoints, adaptation choices and budgets limit architecture-only comparisons. Table 3 separately uses 2,000-step Psi0 fine-tuning. e10e16
During inference
Policy servers receive RGB/proprioception, return action chunks, and clients queue commands. MuJoCo physics is specified at 500 Hz; the Gym loop and rendering sequence are described at 50 Hz. These nominal rates do not establish real-time throughput, given the separately reported approximately four rendered frames per second. e03e13e14
AMO decouples upper- and lower-body tracking; SONIC offers whole-body tracking. Learned action prediction and controller execution remain separate. DreamZero is the evaluated model that jointly predicts video and actions; SIMPLE supplies the physical world and evaluation interface. e16e17
5.5 Implementation flow
- Prepare physical and visual assets
CoACD creates collision geometry, while USD conversion supplies textured rendering assets. Free-drop simulation finds stable object poses; BoDex caches feasible grasps, which CuRobo later uses for scripted arm trajectories. This avoids online grasp synthesis. e04
- Collect behavior before expensive rendering
Motion planning decomposes tasks into manipulation and navigation. Alternatively, PICO teleoperation maps hand movements through inverse kinematics while a whole-body controller handles balance and joystick navigation. Isaac Sim is disabled during VR collection; MuJoCo stereo views provide feedback and demonstrations are recorded at 50 Hz. e05
- Replay and diversify observations
Trajectories are replayed in Isaac Sim with changes to objects, poses, materials, lights, viewpoints and instructions. Figure 2 includes success-only filtering. Its Level 0/1/2 progression adds distractors, appearance changes and spatial changes; Section 3.5 instead labels the corresponding progression 1/2/3. e03e06
- Connect policy predictions to execution
The 32 state coordinates cover hands, arms, waist rotation and base height. Policies output 32 corresponding targets plus four navigation commands. Whole-body controllers translate them into joint commands, MuJoCo advances contact dynamics, and synchronized Isaac Sim images supply subsequent visual feedback. e03e07e14e17
6. Experiments & results
SIMPLE combines MuJoCo physics, Isaac Sim rendering, demonstration collection and a common humanoid policy interface. Its central contribution is an evaluation and training environment. Experiments show task-dependent policy rankings, benefits from visual diversity and more demonstrations, and simulation-trained Psi0 achieving 8/10 physical successes on each of two tasks under similar settings. These findings support a useful testbed, while leaving broad real-world ranking fidelity incompletely established.
6.1 Read the original evidence
Table 1. Policy performance depends strongly on the task and embodiment adaptation. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Choose a task column before comparing model rows. Each slash-separated cell follows Level 0 / Level 1 / Level 2, as specified by the caption; the slashes do not denote a fraction. Bold and underlined entries mark best and second-best performance. The selected tasks mix teleoperation and motion-planning demonstrations, and the experiments use single-task fine-tuning. Inspect XMoveBendPick to see ACT beside Psi0, then inspect XMovePick to see DreamZero's different profile. The table reports performance entries between zero and ten, but its caption does not explicitly define the denominator, so the original triplets are preserved without converting them to percentages. e08e16
What it supports. ACT reaches 9 / 10 / 10 on XMoveBendPick, compared with Psi0's 10 / 9 / 9 and DreamZero's 0 / 0 / 1. DreamZero instead reaches 10 / 10 / 10 on XMovePick. This contrast supports task-dependent competence, rather than one model family dominating every evaluated behavior.
Where the evidence stops. These are six selected tasks, not the entire advertised suite. Checkpoints, training budgets, action horizons and adaptation choices differ across baselines. The table therefore cannot isolate the causal value of model size, video prediction or pretraining.
Table 2. Simulation teleoperation improves collection throughput in the reported sessions. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each task has two columns: demonstrations per hour, where higher is better, and average episode duration in seconds, where lower is better. Compare rows within the same task before comparing tasks of different complexity. Motion Planning (Sim) generates behavior autonomously; both teleoperation rows use one experienced operator per session. The three task groups are whole-body pick-place, stand-still handover and mobile pick-place. The simulation teleoperation row leads on the displayed speed measures for all three. These measurements concern the demonstration collection stage illustrated on the left of the pipeline; photorealistic replay is a subsequent processing stage. e03e09e13
What it supports. For mobile pick-place, simulation teleoperation gives 156.5 demonstrations per hour at 23.0 seconds per episode, versus 87.2 and 41.3 for real teleoperation. Planning gives 24.0 and 150.0, but runs without an operator. The table exposes a tradeoff between human-assisted collection speed and autonomous generation.
Where the evidence stops. The source does not provide operator-to-operator variability or a complete wall-clock accounting through rendering. Its separate limitation reports approximately four ray-traced frames per second per GPU, so collection throughput should not be read as finished-dataset throughput.
Table 5. Simulation-only downstream adaptation transfers to two physical tasks under similar settings. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row horizontally to compare a task in simulation and on the physical robot. The caption describes a single policy fine-tuned exclusively on simulation data; Section 4.4 identifies it as Psi0. Pick & Place falls from nine simulated successes to eight physical successes, while Handover falls from ten to eight. The ten-trial denominators are explicit. Figure 5 on the same inspected page supplies paired egocentric simulation and real observations, and the text says environmental settings are similar. Thus, the comparison evaluates downstream sim-to-real transfer in selected settings, rather than a cross-model ranking study. e08e12e15
What it supports. The physical policy succeeds in 8/10 trials on each task without downstream real-world fine-tuning. The simulation-to-physical gaps are ten percentage points for Pick & Place and twenty for Handover. This is direct execution evidence on two tasks, complementing the larger set of simulation-only comparisons.
Where the evidence stops. Zero-shot refers to transfer after simulation fine-tuning, not absence of pretrained knowledge. Two tasks and ten trials per task do not establish robustness to unseen environments or the paper's broader claim that simulation preserves rankings across policies.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| XMoveBendPick policy comparison Single-task adaptation; Table 1 Level 0 / 1 / 2 simulation settings. | ACT: 9 / 10 / 10. Reported performance triplet; Table 1 does not explicitly define its denominator. | Psi0: 10 / 9 / 9; DreamZero: 0 / 0 / 1; pi0.5: 0 / 0 / 0. ACT is competitive on this coordinated movement task; the result does not isolate model size or world-model pretraining. e08e16 |
| Mobile pick-and-place demonstration collection Table 2 T3; one experienced teleoperator per teleoperation session. | Simulation teleoperation: 156.5 demos/hour; 23.0 seconds. Demonstrations/hour and mean episode seconds. | Real teleoperation: 87.2; 41.3 seconds. Automated simulation planning: 24.0; 150.0 seconds. Simulation reduces collection time here; planning needs no operator. These figures do not measure the complete rendered-dataset pipeline. e09e13 |
| BendHandover randomization ablation Psi0 fine-tuned for 2,000 steps; ten training examples, either all Level 0 or five each from Levels 0 and 1. | Mixed training: 8/10, 7/10, 8/10 on Sets 0, 1, 2. Successes in ten evaluations per set. | Level-0-only: 8/10, 8/10, 5/10. Set 2 improves by 30 percentage points; Set 1 declines by ten points. This supports a targeted robustness gain, without uncertainty estimates. e10 |
| XmoveBendPick demonstration scaling Psi0; 2,000 fine-tuning steps, 10 versus 100 teleoperation trajectories. | 100 trajectories: 10/10, 9/10, 9/10. Successes in ten evaluations per set. | 10 trajectories: 5/10, 6/10, 3/10. More trajectories improve all three sets here; the table does not establish a general scaling law. e10 |
| Motion planning versus teleoperation training Psi0 on BendPick, Mobile P&P and XMoveBendPick; Table 4. | Teleoperation average: 7.56. Reported average performance and per-level triplets. | Planning average: 5.00; BendPick planning scores 10/10/10 versus teleoperation 8/8/6. The overall teleoperation advantage has a task-specific exception. Unreported size matching prevents clean attribution to motion naturalness. e11 |
| Zero-shot real-robot pick-and-place and handover One Psi0 policy fine-tuned only on simulation data; physical evaluation under similar environmental settings. | Physical Pick & Place: 8/10; physical Handover: 8/10. Task successes in ten trials. | Simulation: 9/10 and 10/10, respectively. Shows downstream transfer without real-world fine-tuning on these two tasks. It does not imply no real-data pretraining or broad environment generalization. e12e15 |
| Extended articulated-object evaluation Psi0, Table 7; Level 0 / 1 / 2. | OpenFaucet: 3 / 3 / 4. Success counts out of ten per level. | CloseDoor and PushOfficeChair: 10 / 10 / 10; OpenTrashCan: 7 / 9 / 9. Difficulty varies substantially. Six additional tasks cannot validate every advertised task; OpenTrashCan also contradicts the prose claim of degradation with randomization. e19 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. Visual diversity and trajectory count test different routes to better generalization. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The horizontal divider separates two experiments, both using Psi0 fine-tuned for 2,000 steps. In the upper block, BendHandover keeps ten training examples while replacing half of the Level-0 data with Level-1 data. Follow the Eval Set 2 column to test harder-setting performance, then check Sets 0 and 1 for a tradeoff. In the lower block, XmoveBendPick changes trajectory count from ten to one hundred. Every evaluation cell explicitly gives successes out of ten and the corresponding rate. Keep the two blocks distinct: one changes training composition, while the other increases the amount of demonstrated behavior. e06e10
What it supports. For BendHandover, the mixed set raises Eval Set 2 from 5/10 to 8/10 while leaving Set 0 at 8/10; Set 1 drops from 8/10 to 7/10. Separately, one hundred XmoveBendPick trajectories yield 10/10, 9/10 and 9/10, compared with 5/10, 6/10 and 3/10 from ten trajectories.
Where the evidence stops. Each cell contains only ten evaluations, with no seed variation or confidence interval. The experiment supports these observed changes; it does not establish universal necessity of randomization or a general data-scaling law. Level numbering also differs elsewhere in the source.
7. Analysis & limitations
7.1 What the evidence leaves open
Rigid-body assumptions exclude faithful deformable-object behavior. Ray tracing is expensive: approximately four frames/second per GPU. The 50 Hz rendering description and differing 500/50 Hz controller statements remain insufficiently reconciled. e13e14e17
The claimed sim-to-real ranking correlation relies on qualitative agreement with prior work, without a coefficient or matched real-world comparison across the nine policies. Physical transfer uses only two tasks with ten trials each and similar scenes. e08e12
GraspNet asset counts are 53 versus 75 in different sections. Figure 9 category counts total 82, although the text advertises 60 tasks; the counting unit is unexplained. Table 6 lists representative tasks rather than a complete inventory. These discrepancies remain unresolved. e04e18e20
7.2 Questions for discussion
- How stable are policy rankings when adaptation budgets and action horizons are controlled?
- Which scene or contact changes first break the demonstrated physical transfer?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires matched MuJoCo/USD robot and object assets, CoACD and cached BoDex grasps, CuRobo, a specified controller, rendering configuration and exact baseline checkpoint/adaptation. Psi0 training hardware and hyperparameters are documented; simulator software versions and detailed randomization ranges are not supplied. e04e06e14e15
Record exact splits, seeds, task thresholds, episode limits and demonstration allocation. Table 6 specifies a greater-than-5-cm lifting condition, but leaves several target angles/distances numeric-free. The paper promises code release; that alone does not establish usable release availability. e02e08e18
Proposed checks should repeat the fixed-size BendHandover randomization comparison across seeds and test whether simulation rankings predict physical rankings under matched scenes and explicit shifts. Neither check has been run. e10e12
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Repeat the fixed-size randomization intervention
Reader-proposed, not run: use the same ten underlying BendHandover demonstrations to construct a Level-0-only training set and a set with five examples rendered at Level 0 and five at Level 1. Keep the Psi0 checkpoint, controller, optimizer and 2,000-step budget fixed. Repeat the assignment and training across multiple seeds, with held-out layouts shared between conditions and at least 50 evaluations per set. Report uncertainty and failures by task phase. A repeatable Set-2 advantage without a comparable Set-0 loss would support the proposed robustness mechanism; disappearance of the gain after matching trajectories would weaken that interpretation. e06e10e15e18
Check 2: Measure whether simulated rankings survive physical shifts
Reader-proposed, not run: compare Psi0, ACT and DreamZero on the two transfer tasks using matched demonstration sets, the same controller and explicit per-policy adaptation records. Evaluate at least 30 trials per policy and task in matched simulated/physical scenes, then repeat with controlled lighting and initial-pose shifts. Preserve task success definitions and report trial uncertainty, rank agreement and failure modes separately for each condition. Consistent simulated ordering that predicts physical ordering would support benchmark fidelity. Rank reversals under small shifts would expose the boundary of the current similar-setting transfer evidence, even if one policy retains high absolute success. e08e12e16e17e18
8.3 Reading coverage
Visual audit: All 18 supplied PDF pages were rendered and visually inspected, including title/authors/version, Figures 1–9, Tables 1–7, Algorithm 1, references and embedded supplementary material. The five final crops were separately inspected and retain complete diagrams/table headers and values. Architecture flow was checked against captions and method text. Evidence pages include the appendix training hardware, action/controller details, success criteria and inconsistent task statistics. The randomization-label discrepancy and unreconciled timing claims are preserved. Separate supplements, external code and cited papers remain outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title, Abstract and Section 1 Introduction
- Sections 2.1–2.3 Related Work
- Sections 3.1–3.5 Method and Algorithm 1
- Sections 4.1–4.4 Experiments
- Section 5 Conclusion and Limitations
- References and supplementary contents
- Sections 5.1.1–5.1.2 System Architecture Details
- Section 5.2 Offline Object Preprocessing Pipeline
- Section 5.3 Baseline Implementation Details
- Section 5.4 Whole-Body Controller Details
- Sections 5.5–5.6 Extended Results and Dataset Statistics
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The complete supplied 18-page text and all PDF pages were inspected, including embedded supplementary material; PDF inspection resolves the figure-extraction limitation above. Separate supplements remain unverified.
- Only arXiv:2606.08278v1 dated 6 June 2026 was supplied. Its title and all twelve authors match the catalog; no other edition was compared. The catalog affiliation contains an extraction artifact; metadata retains only the observed title-page affiliation.
- Code, external project resources and cited papers were not inspected. No experiments were run.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block and arXiv margin
The title is SIMPLE: Simulation-Based Policy Learning and Evaluation for Humanoid Loco-manipulation. The margin identifies arXiv:2606.08278v1 [cs.RO], 6 Jun 2026. Authors: Songlin Wei, Zhenhao Ni, Jie Liu, Zhenyu Zhao, Junjie Ye, Hongyi Jing, Junkai Xia, Xiawei Liu, Michael Leong, Liang Heng, Di Huang, Yue Wang. Affiliation: USC Physical Superintelligence (PSI) Lab. The first four authors contributed equally; Yue Wang is corresponding author.
Go to primary source ↓e02PDF pp. 1–2, Abstract, Figure 1 caption and Section 1
SIMPLE addresses costly, difficult-to-standardize humanoid evaluation through simulation, demonstration collection and policy benchmarking. The introduction reports 60 tasks, 50 scenes, over 1,000 objects and over 6,000 trajectories. The abstract continuation promises future code release.
Go to primary source ↓e03PDF p. 4, Figures 2–3, Section 3.1 and Algorithm 1
MuJoCo handles physics/control and sends states to Isaac Sim, which sends rendered images to the policy; actions return to MuJoCo. Figure 2 shows planning and teleoperation, offline replay, success-only filtering and evaluation randomization. Algorithm 1 labels a 50 Hz control loop.
Go to primary source ↓e04PDF p. 5, Section 3.2; p. 15, Figure 8 and Section 5.2
CoACD collision decomposition and USD conversion prepare assets. Free-drop stable poses and BoDex grasps are cached for CuRobo. Both sections report more than 1,500 Objaverse objects, but GraspNet counts differ: 53 on p. 5 versus 75 on p. 15. Section 3.2 specifies 50 HSSD scenes.
Go to primary source ↓e05PDF p. 5, Figure 4 and Section 3.3
Scripted decomposition combines CuRobo dual-arm planning with lower-body navigation. VR collection disables Isaac Sim, streams MuJoCo stereo views to PICO, retargets hand poses through inverse kinematics, and uses joystick navigation and whole-body tracking. Demonstrations are recorded at 50 Hz in LeRobot format.
Go to primary source ↓e06PDF p. 4, Figure 2; p. 5, Section 3.4; p. 6, Section 3.5; p. 14, DomainRandomizer paragraph
Replay randomizes objects, poses, textures, lights, viewpoints and instructions. Figure 2 and the appendix define Level 0 as backgrounds/distractors, Level 1 as additionally materials/lighting, and Level 2 as additionally object/robot poses. Section 3.5 instead numbers the corresponding progression 1, 2, 3.
Go to primary source ↓e07PDF p. 6, Section 3.5, whole-body representation paragraph
State dimension is 32: 14 hand, 14 arm, three waist rotation coordinates and base height. Actions add four locomotion commands, listed as v_x, v_y, v_yaw and q_yaw, to those 32 targets. The text reports 6,000 rendered episodes.
Go to primary source ↓e08PDF p. 6, Table 1, caption and Section 4.1; p. 7, Experimental Results
Nine baselines are compared on six tasks using single-task fine-tuning. Table 1 orders entries Level 0 / 1 / 2 and calls them performance without explicitly defining a denominator. XMoveBendPick: ACT 9/10/10, Psi0 10/9/9, DreamZero 0/0/1, pi0.5 0/0/0. XMovePick: DreamZero 10/10/10, Psi0 10/10/6. Bold/underline denote best/second-best. The prose asserts agreement with prior real-world rankings without reporting a correlation coefficient or paired multi-policy physical results here.
Go to primary source ↓e09PDF p. 7, Table 2 and Section 4.2
Whole-body pick-place real/sim teleoperation: 206.8/310.3 demos per hour, 17.4/11.6 seconds. Handover: 130.9/197.8, 27.5/18.2. Mobile pick-place: 87.2/156.5, 41.3/23.0. Planning: 58.9, 32.7, 24.0 demos per hour and 61.1, 109.8, 150.0 seconds. Planning is operator-free; teleoperation uses one experienced operator per session.
Go to primary source ↓e10PDF p. 7, Section 4.3; p. 8, Table 3
Psi0 is fine-tuned for 2,000 steps. BendHandover with 10 Level-0 examples scores 8/10, 8/10, 5/10 on Sets 0–2; a 5+5 Level-0/1 mixture scores 8/10, 7/10, 8/10. XmoveBendPick with 10 teleoperation trajectories scores 5/10, 6/10, 3/10; with 100 it scores 10/10, 9/10, 9/10.
Go to primary source ↓e11PDF p. 8, Table 4 and Data Source Quality paragraph
Planning-only Psi0: BendPick 10/10/10, Mobile P&P 3/2/2, XMoveBendPick 4/2/2, Avg. 5.00. Teleoperation-only: 8/8/6, 7/5/6, 10/9/9, Avg. 7.56. The authors attribute the overall advantage to diverse, naturalistic motions; dataset-size matching is not stated.
Go to primary source ↓e12PDF p. 8, Table 5, Section 4.4 and Figure 5
A Psi0 policy fine-tuned exclusively on simulation data is tested in similar simulation/physical settings without downstream real-world fine-tuning. Pick & Place: 9/10 simulated, 8/10 physical. Handover: 10/10 simulated, 8/10 physical. Figure 5 shows adjacent simulation/real egocentric pairs for these tasks.
Go to primary source ↓e13PDF p. 9, Limitations paragraph
Ray tracing achieves approximately four frames per second on one GPU, whose model is not specified here. Rigid-body assumptions prevent faithful simulation of cloth, rope and other deformable/soft objects.
Go to primary source ↓e14PDF pp. 13–15, Figures 6–7 and Sections 5.1.1–5.1.2
The Gym environment owns physics/rendering backends and a task; reset configures a shared layout. MuJoCo physics is 500 Hz. Clients send RGB/proprioception to separate policy servers, make blocking calls and enqueue action chunks. Figure 7 and Section 5.1.2 describe RGB rendering at 50 Hz without reconciling wall-clock throughput with p. 9.
Go to primary source ↓e15PDF p. 15, Section 5.3, Psi0 paragraph; p. 16, opening continuation
Psi0 has a Qwen3-VL-2B vision backbone and flow-matching transformer action expert. Its released checkpoint is fine-tuned for 40,000 steps on eight NVIDIA A100 GPUs with DDP, batch 16 per device, cosine learning rate 0.0001, 1,000 warmup steps and bf16. Output dimension is 36; action chunk size is 30.
Go to primary source ↓e16PDF p. 16, Section 5.3, DreamZero through ACT paragraphs
DreamZero uses 14B Wan2.1-I2V and the DreamZero-AgiBot checkpoint: rank-4 LoRA, 40,000 steps, eight A100s, ZeRO-2, batch 1 per device, learning rate 0.0001, 5% warmup and action horizon 48. pi0.5 expands 30 actions to 36 via padded weights. GR00T trains 20,000 steps on three A100s and uses sequential inference because RTC code is unavailable. InternVLA freezes its VLM and trains its action head; EgoVLA adapts to joint commands. DP uses 100 denoising steps. ACT has 36 action outputs, chunk size 100, four encoder layers and one decoder; other hyperparameters follow DP.
Go to primary source ↓e17PDF p. 15, Whole-Body Controllers paragraph; pp. 17–18, Section 5.4
AMO separates upper-body tracking from a learned lower-body controller; SONIC tracks whole-body motion, with direct and decoupled implementations. The general p. 17 controller description says outputs reach G1 at 500 Hz; the SONIC-specific p. 18 continuation and p. 15 specify 50 Hz. These rates are not fully reconciled.
Go to primary source ↓e18PDF p. 17, Table 6, Task Name and Success Criteria columns
The table lists representative tasks, not a complete 60-task enumeration. BendPick, BendPickMP and TableTopGraspMp require a lift exceeding 5 cm. Handover requires transferring the object and placing it in a container. Several articulated tasks refer to target angles/distances without numeric thresholds.
Go to primary source ↓e19PDF p. 18, Section 5.5 and Table 7
Additional Psi0 results: CloseDoor 10/10/10, OpenOven 7/5/4, OpenFaucet 3/3/4, P&P&HugContainer 7/6/3, PushOfficeChair 10/10/10, OpenTrashCan 7/9/9. The text treats six additional tasks as validating every task and describes articulated-task degradation despite OpenTrashCan increasing.
Go to primary source ↓e20PDF p. 18, Figure 9 legend and Section 5.6
The legend gives Basic (33), Articulated (5) and Loco-manip. (44), with 40%, 6%, 54%. These counts total 82, whereas Section 5.6 describes 60 tasks; the difference is unexplained.
Go to primary source ↓8.5 Primary sources
SIMPLE: Simulation-Based Policy Learning and Evaluation for Humanoid Loco-manipulation ↗
PDF · 9,191 extracted words
Source fingerprint
f9150c3c8db27b0c9a2765b242ca2cc8f837d63dedd43b46b68dc00cc42d72c9