Inverse kinematics (IK) converts end-effector pose commands into joint motion, and many vision-language-action (VLA) policies rely on it to act on a given arm and to transfer across embodiments. Precise IK remains laborious: redundancy and singularities complicate the inverse mapping, execution depends on hand-tuned controllers, and training data on physical robots are costly to collect. We present PolicyIK, a goal-conditioned reinforcement learning policy for high-precision IK. It updates joint-position references from pose feedback and observes its accumulated reference relative to the measured joints, so it can retain the offset that the position controller needs at rest. Two-scale pose errors with an error-dependent action gain resolve millimeter residuals while preserving large motions. Progress and dwell rewards teach it to both reach the tolerance and stay within it. In simulation, a single attempt by the trained PPO policy keeps the mean errors over the final second within 1 mm and 0.2° on 93.5% of held-out tasks, and a deployment protocol raises success to 98.9%. When trained directly on a physical arm with an off-policy RL learner, PolicyIK achieves higher reaching and holding success than all four classical IK controllers under the same time budgets. In simulated pick-and-place, PolicyIK executes the end-effector commands of the π0.5 VLA policy with 95.7% success. These results show that reinforcement learning can provide high-precision inverse kinematics for robot arms. Code is available at https://github.com/Tinaliu0123/PolicyIK.
Figures & tables
Figure 1: PolicyIK: from broad reaching to precise regulation. (a) Full-pose reaching across the simulated workspace. (b) Reaching and holding within 1 mm and 0.2∘ . (c) On hardware, PolicyIK holds the pose within 1 mm and 0.2∘ throughout the final 5 s on more tasks than classical controllers.
Figure 2: PolicyIK control formulation. At each step, the policy maps the observation to a joint-reference increment, which is scaled by the error-dependent gain, accumulated in the joint reference, and tracked by a joint PD controller. Colored blocks are the designs of Sections 4.1 to 4.3 .
Median error
Controller
Strict (%) ↑
Coarse (%) ↑
Pos. (mm) ↓
Ori. ( ∘ ) ↓
E[N]
UMI-RL (tuned)
0.00 ± 0.00
16.38 ± 3.10
16.65 ± 0.80
4.25 ± 0.77
1.000
DLS
50.41
64.42
0.94
0.15
1.000
QP
56.21
63.98
0.57
0.11
1.000
MPC
56.33
64.78
0.54
0.10
1.000
PolicyIK
93.49 ± 0.20
96.56 ± 0.16
0.25 ± 0.02
0.04 ± 0.00
1.000
Table 1: Simulation results. With deployment, each task reports its first strictly successful attempt, or its last attempt if none succeeds, and E[N] is the mean number of attempts. DLS, QP, and MPC are tuned on a separate set of tasks from the same distribution (Appendix E ).
Figure 3: Ablations on command-state feedback, pose feedback and gain, and rewards. Curves show the share of test tasks within 1 mm and 0.2∘ at each time step.
Figure 5
Table 3: Reaching, holding, and median pose errors on 60 physical tasks within 25 s. The best result is in bold, and the second-best is underlined.
Figure 6: A case in which MPC jitters around the goal after reaching, while PolicyIK holds.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Strict (%) ↑
E[N]
Final checkpoint, single attempt
92.30 ± 0.13
1.000
Checkpoint average, single attempt
93.49 ± 0.20
1.000
+ error-corrected goal
94.80 ± 0.12
1.036
+ transient-goal attempts
98.92 ± 0.23
1.212
Appendix
Table 4: Strict success as checkpoint averaging and the two deployment steps are added. E[N] is the mean number of attempts per task.
Median error ↓
Setting
Coarse (%) ↑
Pos. (mm)
Ori. ( ∘ )
0.20 s horizon, k=1 (source)
0.04 ± 0.01
19.45 ± 3.80
74.69 ± 2.62
0.54 s horizon, k=1
0.02 ± 0.02
6.27 ± 0.43
54.87 ± 3.85
0.54 s horizon, k=0.5
0.35 ± 0.09
11.17 ± 2.27
37.45 ± 0.61
0.54 s horizon, k=0.25
16.38 ± 3.10
16.65 ± 0.80
4.25 ± 0.77
3.3 s horizon, k=0.1
1.04 ± 0.06
44.16 ± 3.52
4.20 ± 1.25
Appendix
Table 5: Tuned variants of UMI-RL. Entries are means and standard deviations across three training seeds. Position and orientation errors are median terminal-window errors. The factor k scales the orientation ladder of the source recipe. Strict success is zero in every setting.
Median error ↓
Controller
Strict (%) ↑
Coarse (%) ↑
Pos. (mm)
Ori. ( ∘ )
DLS, selected ( k=2 , 0.15 rad lead)
50.41
64.42
0.94
0.15
QP, selected ( k=1.5 , 0.15 rad lead)
56.21
63.98
0.57
0.11
MPC, selected ( k=1.5 , 0.15 rad lead)
56.33
64.78
0.54
0.10
DLS, measured-joint anchor, full step
1.84
46.50
4.74
2.20
DLS (default), 30 Hz
1.60
42.86
5.88
2.73
Appendix
Table 6: Classical controllers on the 4,120 test tasks. The selected settings appear in Table 1 . DLS (default) is the Isaac Lab differential IK controller with its default settings, as in Section 5.1.3 .
Figure 7: One test task. Left: intervals in which each joint lies within 0.01 rad of its lower (dark) or upper (light) limit. Right: position and orientation errors, with the strict tolerances dashed. The dotted line marks when a third QP joint reaches its limit (1.7 s). The QP failure coincides with this sustained limit occupancy, and its error then stays near 96 mm and 8∘ . PolicyIK never reaches a joint limit and enters the strict tolerance at 2.0 s.
Success (%) ↑
Terminal phase ↓
Median error ↓
Variant
Coarse
Strict
Entry (s)
Uncont. (%)
Pos. (mm)
Ori. ( ∘ )
PolicyIK
96.56 ± 0.16
93.49 ± 0.20
1.73 ± 0.01
2.43 ± 0.23
0.25 ± 0.02
0.04 ± 0.00
Command-state feedback and reference anchor
w/o command-state feedback
92.63 ± 0.20
84.26 ± 0.28
2.04 ± 0.02
10.32 ± 0.32
0.30 ± 0.00
0.06 ± 0.00
measured-joint anchor
95.90 ± 0.17
30.77 ± 7.94
1.80 ± 0.03
82.46 ± 6.62
1.27 ± 0.05
0.16 ± 0.01
Pose feedback and action scaling
Appendix
Table 7: Component ablations with one attempt per task and matched training budgets. Entry is the median first-entry time and Uncont. the uncontained share (Appendix C ). The measured-joint anchor sets zt−1=qt in Equation 2 and removes command-state feedback, and the fixed gain is β=0.1 .
Targets
DLS (default)
DLS (hand-tuned)
QP (hand-tuned)
PolicyIK
Closest to the mount
85.5
96.8
85.9
95.1
Farther to the side
96.1
96.5
95.7
96.4
Appendix
Table 8: Pick-and-place success (%) by target location, three runs from each of the 50 official initial states per task (750 episodes per group).
Figure 8: (a) The FlashSAC network, used for the actor and, in FlashSAC, for each critic. Filled circles mark the hyperspherical normalization ( Lee et al., 2025 ) , where weight vectors project onto the unit sphere after every gradient step. (b) The latent-alignment critic that PolicyIK uses on hardware (Appendix H.2 ). In both panels, pω(ot,at) denotes the value-distribution logits that the critic’s linear head outputs; a softmax over the atoms turns them into the probabilities pω,i of Equation 15 .
15 s budget
25 s budget
Policy
Reached
Held 1 s
Held 5 s
Reached
Held 1 s
Held 5 s
Coarse
Gross
Ablated
40
36
31
46
43
40
52
5
PolicyIK
52
44
27
54
45
43
59
1
Appendix
Table 9: Ablation of the hardware components on 60 physical OpenArm tasks. The Ablated policy excludes the latent-alignment critic, far-task selection by joint-space distance, and the relaxed reference bound. Reached and held use the criteria defined in the main body. Coarse counts tasks whose final-second errors are within 1 cm and 2∘ , and gross counts tasks that end more than 20 mm from the goal, both at the 25 s budget. Bold marks the better result.
15 s budget
25 s budget
Controller
Reached
Held 1 s
Held 5 s
Reached
Held 1 s
Held 5 s
DLS
38
21
6
48
29
17
Resolved-rate QP
43
21
6
48
24
16
Kinematic MPC
29
10
3
46
26
15
Adapted OpenArm QP
13
8
1
32
16
11
PolicyIK
52
44
27
54
45
43
Appendix
Table 10: Reaching and final-window holding counts out of 60 OpenArm tasks, with best results in bold and second-best results underlined.
Setting
PolicyIK on hardware
FlashSAC
Learner
Online transitions
120,000
Batch size
1,024
2,048
Replay capacity
120,000 (the whole run)
106
Initial random actions
200
–
Updates begin after
1,024 transitions
–
Appendix
Table 11: Settings for training PolicyIK on hardware, used for both the OpenArm and the xArm7. The FlashSAC column lists the corresponding settings of FlashSAC ( Kim et al., 2026 ) ; a dash marks a setting that FlashSAC does not specify.
Controller
Formulation
Parameters
DLS
Damped least squares with global speed scaling
k=2 , λ=0.02
Resolved-rate QP
Damped least squares within the joint bounds
k=3 , λ=0.005
Kinematic MPC
Seven-step receding horizon (0.14 s)
k=3 , λ=0.005
Adapted OpenArm QP
OpenArm QP with measured-pose settling
Costs 12/1.5, damping 0.1
Appendix
Table 12: Classical hardware baselines. All four map the measured joint state and the goal pose to joint-position references at 50 Hz within the joint limits and the 0.20 rad/s reference-speed bound. DLS, QP, and MPC weight rotation errors by 0.2 relative to translation.
Controller
Success
Reach (s)
Position (mm)
Orientation ( ∘ )
Standard reaching
DLS
60/60
2.9
0.074
0.015
QP
60/60
2.0
0.094
0.018
MPC
60/60
2.3
0.085
0.016
xArm built-in IK
59/60
1.8
0.004
0.001
PolicyIK (sim-trained)
60/60
2.2
0.169
0.034
Appendix
Table 13: Physical xArm7 evaluation on two sets of 60 tasks each. Reach-time medians use reached tasks.
Inverse kinematics maps desired robot or articulated-body motions to joint configurations while handling constraints, redundancy, and multiple solutions.
Inverse kinematics (IK) remains a critical bottleneck for real-time robot manipulation. Classical numerical solvers achieve high geometric precision but often suffer from discontinuous branch switching and unstable behavior near kinematic singularities during closed-loop deployment. Meanwhile, learned IK approaches frequently struggle to balance spatial accuracy, motion smoothness, and real-time efficiency, particularly when trained on noisy human teleoperation data. We present MimicIK, a real-time generative inverse kinematics framework that learns smooth and robust joint-space motion priors from teleoperation demonstrations through conditional flow matching. Given the current joint configuration and a target end-effector pose, MimicIK predicts continuous delta-joint commands using an efficient two-step iterative refinement process based on a Minimal Iterative Policy (MIP) backbone. To enforce physical consistency, we further introduce an FK consistency loss, a differentiable forward-kinematics regularization that penalizes task-space deviations from the target pose during training. We evaluate MimicIK on a real-world 6-DOF robot dataset containing 8,848 teleoperation demonstrations. MimicIK achieves a mean position error of 4.65 mm, a 10 mm success rate of 92.01%, and a trajectory spike rate of only 7.99%. Compared with a UNet diffusion baseline, our method improves both spatial accuracy and motion smoothness while reducing inference latency from 21.66 ms to 6.74 ms. Furthermore, unlike deterministic MLP baselines that catastrophically diverge under out-of-distribution deployment, MimicIK remains stable near singular configurations and enables robust 20 Hz real-time control on deployment hardware.
Neural models can learn to generate various solutions to the inverse kinematics problem from data, but are usually limited to a single robot. We present MorphIK, a flow-matching model that solves inverse kinematics for revolute-joint-based kinematic chains it has never seen during training. The model uses a transformer architecture to encode the robot's morphology along with the target pose. This encoding then conditions a flow-matching head that generates poses from noise. Trained on purely synthetic data from procedurally generated robots, the model reaches a precision of about 5 cm on unseen real-world robots with 6 to 9 Degrees of Freedom. For higher precision, the model serves as an excellent Prior for further optimization algorithms, reducing error to less than 1 cm after a single step of Damped Least Squares optimization and to sub-1 mm error after 3 steps in most cases. Building on flow matching's generative capabilities to produce highly diverse outputs, our model can efficiently sample the robot's null space, providing a wide variety of configurations for the same pose. Thus, overall, MorphIK allows learning and generalizing neural inverse kinematics for a multitude of known and unknown robots.
Lennart Clasmeier, Jan Gerrit Habekost, Cornelius Weber +1
Knowledge Technology Group, Department of Informatics, University of Hamburg, Hamburg, Germany
Robotic manipulation robustness often founders on the physics gap between simplified simulations and the resistance-laden real world. In this work, we emphasize that physical realism in articulated interaction is an important ingredient for robust policy learning. We present Real-IKEA, a dataset and simulation framework designed with physical accuracy as a first-class goal. Real-IKEA provides 1,079 articulated asset configurations, derived from 83 authentic IKEA handles and knobs processed through a meticulous six-step physical workflow. For contact-geometry accuracy, we introduce a bidirectional surface-deviation metric to quantify collision meshes. For dynamics realism, we establish resistance-calibrated configurations that vary damping and friction. Crucially, we demonstrate through a Reinforcement Learning (RL) policy that high-fidelity assets enable the discovery of robust "hooking" and "levering" strategies that prioritize mechanical advantage over fragile friction-pulling. Together, these results position Real-IKEA as a critical benchmark for developing manipulation policies capable of human-level robustness in articulated object tasks.
Kunqi Xu, Zhenhao Huang, Siyuan Luo +2
Human-centered Robotic Lab, National University of Singapore · School of EECS, Peking University