We evaluate Genie Sim Benchmark through paired real-simulation tasks and complementary system-identification studies. Across 22 robot-task pairs spanning 16 task names, task-specific pi0.5 policies fine-tuned on real-robot data achieve a descriptive Score correlation of 0.9686, MAE 0.0405, RMSE 0.0520, with simulation-real differences from -0.110 to +0.080. G2 crsB experiments assess long-horizon closed-loop position and joint-side torque agreement, while five representative object cases examine identification and cross-engine contact transfer. A separate 32.6503 s teleoperation trajectory, excluded from mechanical identification and servo-gain search, evaluates external position targets through simulated motion-controller command shaping and a calibrated position servo. Across both arms, per-joint position RMSE spans 0.00614-0.03522 rad and current-derived torque RMSE spans 0.73545-4.00099 N m. Native PhysX adaptations recover three recorded interaction tasks, with case-dependent alignment gains. These results support task-specific simulation evaluation through joint assessment of task outcomes and physical-response consistency.
Figures & tables
Task family
Tasks
G1 Omnipicker
G2 Swiftpicker
G2 Wuji
Arx X5
Selection & reasoning
5
3
5
0
3
Placement & sorting
5
4
0
1
0
Bimanual handover
1
1
0
0
0
Stacking
1
0
0
0
1
Textile manipulation
4
0
1
0
3
TABLE 1: Real2Sim evaluation coverage for the four robot configurations illustrated in Figure 2 . Tasks are distinct evaluated task names; robot columns count completed tasks for each configuration. Figure 2 presents representative paired scenes.
Fig. 2: Representative scenes from the Real2Sim experiments summarized in Table 1 : sixteen scene pairs across G1 Omnipicker, G2 Swiftpicker, G2 Wuji and Arx X5 configurations. Rows 1 and 3 are real scenes; rows 2 and 4 are their simulation counterparts. Corresponding columns preserve the paired object arrangements and interaction context.
Stack
Version / path
Model route
Interpretation
MuJoCo
3.10.0 CPU
Frozen native G2 + task model
Source fit; soft contact settings
Newton / MJWarp
MuJoCo-Warp 3.10.0 CUDA; Newton 1.4.0
Frozen MJB via SolverMuJoCo
Same contact family; GPU repeat variation
PhysX
Isaac Sim 6.0.0 RC ⋅ CPU TGS
USD built from frozen numeric arrays
Original hard-contact transfer; updated native contact/damping adaptations ( Appendix D )
TABLE 2: Reported execution stacks in the object study.
Fig. 3: Implemented Real2Sim workflow and corresponding evidence. Robot identification, object replay and paired task evaluation provide complementary measurements; arrows organize the process rather than denote a single end-to-end experiment.
Fig. 5: Long-horizon closed-loop consistency with frozen parameters. Top: seven-joint RMSE across three excitation families, with separate position/torque scales. Bottom: absolute-error CDFs, equally weighted by family per joint; logarithmic error axes include the full observed upper tail. Full native rows score each cell; time samples are correlated. Torque is uncentred, and the J2 mismatch remains visible. Separate joint-side sensor diagnostics appear in Appendix C.4 .
Arm
Torque RMS (N ⋅ m)
Scored joint-time rows
Left
0.7709
6,986
Right
0.9698
6,986
TABLE 3: Final-model Fourier torque reconstruction, approximately 100 s. One recording per arm; all seven joints; own-offset RMS. Same-recording diagnostics, including fitting intervals.
Arm
Joint
Armature J a
Coulomb f_c
Viscous f_v
Bias b τ
Left
J1
1.5165
5.2766
4.1386
0.0246
Left
J2
1.5031
5.4307
2.7398
1.1140
Left
J3
0.5690
1.8589
1.2645
0.0177
Left
J4
0.5717
1.6363
1.9437
0.2242
Left
J5
0.2869
1.3225
0.6124
0.0111
Left
J6
0.2871
1.9650
0.6518
0.0568
TABLE 4: Identified joint parameters for both arms, J1–J7. Armature J a : kg ⋅ m 2 ; Coulomb f_c and bias b τ : N ⋅ m; viscous f_v: N ⋅ m ⋅ s/rad. Values retain the supplied four-decimal precision.
Asset
Real observations
Search and adjusted quantities
Kettle
260 RGB-D lid observations
Geometry, COM, yaw, contact; 3 grid rounds
Gum box
127 RGB-D centre observations
Contact; 100 candidates
Sponge
92 RGB-D centre observations
Effective contact; 120 MuJoCo candidates
Shuttlecock
12 RGB points; 3 valid depths
Sliding/rotational contact; 81 candidates
Pen
11 RGB points and valid depths
81 contact + 42 radius candidates; rejected
TABLE 5: Signature benchmark assets, observation coverage and identification scope.
Fig. 6: Fifteen case–stack mean pairs, compacted for the main text. PhysX † uses the matched native configurations in Table D3 ; other rows retain source results. Hollow markers are baseline, filled markers are fitted / candidate models; red marks increased error. The source pen candidate remains unaccepted. Units and log axes differ between panels. Historical repeat ranges and failures remain in Appendix D ; no confidence intervals are inferred.
Fig. 7: Two evaluated object-replay protocols. The source study freezes the MuJoCo candidate for three-stack comparison. The native PhysX experiments apply the same accepted engine settings to both parameter sets and separately declare the sponge friction refit. Both protocols evaluate task completion and observation error. The dashed strip outlines future VLM priors, agent-selected interactions and independent real evaluation; these steps have not been executed. Author-created schematic, not an additional experiment.