Source-linked AI summary
MorphIK: Morphology-Conditioned Neural Inverse Kinematics for Unknown Robots
Lennart Clasmeier, Jan Gerrit Habekost, Cornelius Weber, Stefan Wermter
TL;DR
Neural inverse-kinematics models are usually limited to a single robot, while redundant chains admit multiple solutions. MorphIK conditions a flow-matching model on robot morphology and target pose, then generalizes from procedurally generated training robots to unseen real-world robots. Its outputs can be polished with DLS for higher precision.
Problem
Neural IK models are usually limited to one robot, despite inverse kinematics remaining an open challenge for redundant kinematic chains.
Source: §I · p. 1Method
MorphIK uses a Transformer and flow-matching decoder conditioned on robot morphology and target pose, trained on procedurally generated robots.
Source: §I · p. 1Results
About 4% of reach mean positional error and 3.2° orientation error without further setup; three DLS steps reduce residual error to sub-0.1% of reach and 1° range.
Source: §VI · p. 8Takeaways & Limitations
Generated joint configurations serve as priors for DLS, reducing residual error to sub-0.1% of reach and 1° range after three steps.
Source: §VI · p. 8
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
MorphIK addresses the single-robot limitation of neural IK with a transformer-based solver trained across procedurally generated robots. It generalizes to unseen real-world manipulators, with Damped Least Squares polishing reducing error further.
I. INTRODUCTION
- Neural IK solvers are usually trained for one kinematic chain, making data collection and retraining costly barriers to integrating and adapting them.MorphIK is proposed to generalize to unseen 6- to 9-DoF revolute serial manipulators.Source: §I · p. 1
- A Transformer backbone and flow-matching decoder generate IK solutions conditioned on the target pose, while training across many randomly generated robots.The robot’s URDF is converted into a canonical representation and tokenized as geometry input.Source: §I · p. 1
- Less than 1 cm error after one DLS step and sub-millimeter accuracy after three demonstrate generalization to real-world manipulators unseen during training.Training is self-supervised, using sampled joint configurations and forward kinematics to obtain corresponding poses.Source: §I · p. 2
II. RELATED WORK
Prior IK approaches include Jacobian-based solvers, genetic algorithms, and neural models. MorphIK extends generative neural IK toward multi-robot use by conditioning flow matching on a robot representation.
II. RELATED WORK
- Jacobian-based solvers such as KDL and Trac-IK can encounter issues near singularities; damping and singularity avoidance address but do not remove practical limitations.Genetic algorithms such as Bio-IK avoid reliance on the Jacobian and allow arbitrary constraints, at the cost of higher computational overhead.Source: §II · p. 2
- Neural IK offers fast, parallel GPU inference and generative models can represent multimodal solutions, enabling exploration of a robot’s null space.Neural models can also use different loss functions and generate training data through forward kinematics on random joint configurations.Source: §II · p. 2
- IKDiffuser generalizes to kinematic trees with multiple end effectors but remains a single-robot model trained individually for each robot.IK Flow uses normalizing flows for generative IK and as a seed model for Trac-IK; MimicIK [4] applies flow matching to teleoperation.Source: §II · p. 2
- MorphIK extends flow matching with a robot representation in its conditioning vector, turning it into a multi-robot model.The related work also describes generative graphical IK using a distance-geometric graph representation to incorporate multiple robots.Source: §II · p. 2
III. APPROACH
MorphIK represents robot morphology canonically and trains on procedurally generated kinematic chains. The dataset comparison shows that generated robots are numerically harder than the real arms used for evaluation.
B. Dataset
- Training uses a Cartesian auxiliary loss and online forward-kinematics samples from stored robot descriptions; evaluation includes 800 held-out synthetic robots and 44 real geometries.The real set includes 6- and 7-DoF arms and constructed 8- and 9-DoF chains.Source: §III · p. 2, B. Dataset · p. 3
A. Canonical Robot Representation
- Canonicalization gives each robot a shared-frame representation, making identical robots identical and geometry tokens comparable across joints, robots, and DoF.The representation uses screw axes and normalizes total chain length to one for scale invariance.Source: A. Canonical Robot Representation · pp. 2–3
B. Dataset
- The synthetic generator varies link lengths, joint-axis relationships, wrist layouts, and joint limits to create diverse revolute chains.It includes intersecting axes, parallel and orthogonal axes, and compact two-joint wrists in some 7–9-DoF chains.Source: B. Dataset · p. 3
- Generated robots have roughly five times lower manipulability and twice the condition numbers of real arms, making them numerically harder to solve.These differences are consistent across DoF classes.Source: B. Dataset · pp. 3–4, Table I · p. 3
- Only 42% of generated 6-DoF chains meet a Pieper criterion, versus 86% of real arms, mainly because fewer generated chains have three consecutive parallel axes.The parallel-axis criterion is met by 5% of generated chains and 57% of real arms.Source: B. Dataset · p. 4
C. Model Architecture
MorphIK encodes robot geometry and the target pose with a transformer, then conditions a flow-matching head to generate robot configurations. The head predicts a velocity field from noise and integrates it over repeated passes.
C. Model Architecture
- The architecture conditions flow-based configuration generation on both robot geometry and the target pose rather than a fixed robot assumption.The transformer output supplies the conditioning for the flow-matching head.Source: C. Model Architecture · p. 4
- The transformer encodes one geometry token per joint, a tool-center-point token, and a token for the target end-effector pose.Its output embeddings are flattened into a conditioning vector for the flow-matching head.Source: C. Model Architecture · p. 4
- An MLP flow-matching head uses AdaLN conditioning from the transformer task encoding to turn noise into a velocity field.Robot token sequences are padded and masked to the maximum 9-DoF length.Source: C. Model Architecture · p. 4
D. Losses
MorphIK combines flow-matching and forward-kinematic losses, using estimated endpoint configurations to add Cartesian pose supervision. The forward-kinematic losses are time-gated, and the model is trained end-to-end with the combined objective.
D. Losses
- MorphIK combines flow-matching and forward-kinematic losses in a single training objective.The transformer encoder and flow-matching decoder are trained end-to-end through this objective.Source: D. Losses · pp. 4–5
- The flow-matching loss uses target velocity from random noise x0 to joint configuration q, with mean squared error between predicted and target velocity.The auxiliary cycle loss adapts Habekost et al.'s differentiable forward-kinematics approach to flow matching.Source: D. Losses · pp. 4–5
- A mostly straight predicted velocity field estimates the endpoint configuration in one step, enabling differentiable forward kinematics to compute Cartesian position and orientation losses.The endpoint estimate supplies the configuration used to calculate the estimated end-effector pose and losses.Source: D. Losses · p. 5
- Forward-kinematic losses are applied only during the second half of the flow and scaled by t^2 because early predictions are poor and endpoint estimates matter late.The final loss is the expected weighted sum of the three loss terms.Source: D. Losses · p. 5
E. Training
Training ran for one million steps with a batch size of 2048, using a constant then annealed learning rate and a weighted combination of flow-matching and forward-kinematic losses.
E. Training
- Training used 1,000,000 steps, batch size 2048, and a learning rate annealed from 5.4e-4 after 350,000 steps to 8e-6.Forward-kinematic loss was weighted 6:1 against flow-matching loss, and position loss 5:1 against orientation loss.Source: E. Training · p. 5
- Evaluation used 10 flow-matching steps; single-solution experiments inferred 16 candidates and selected the best for DLS optimization and reporting.Experiments ran on one Nvidia RTX 5080 with 16GB of VRAM.Source: E. Training · p. 5
IV. EXPERIMENTS
The experiments test MorphIK as a raw solver, a fine-tuned single-robot model, and a seed for optimization. Generalization is evaluated on held-out generated robots and unseen real-world geometries.
IV. EXPERIMENTS
- The study evaluates MorphIK as a raw IK solver, a fine-tuned single-robot model, and a seed for optimization methods such as DLS.The foundation-model experiments compare fine-tuning against training the same model from scratch for the same number of steps.Source: §IV · p. 5
- Generalization tests use held-out robots from the training generator and real-world robots unseen during training.For fine-tuning, three real-world robots per DoF class were selected at random.Source: §IV · p. 5
V. RESULTS
DLS polishing substantially improves MorphIK’s real-robot outputs, reaching sub-millimeter precision after three steps in the hand-crafted architectures. Raw performance is similar across real-robot DoF classes, while polishing reduces errors faster for higher-DoF variants.
V. RESULTS
- Raw real-world robot performance is similar across DoF classes, with 6-DoF robots performing slightly worse.The evaluated real-world robots are hand-designed and not sampled from the training distribution, yet their solutions tend to outperform those for random robots.Source: §V · p. 5
- On synthetic robots, raw mean position error is 3.67% of reach and orientation error is about 2.84°; tail errors are about 9.36% and 7.58°, respectively.Synthetic 6- and 7-DoF robots perform similarly, with higher-DoF robots slightly worse.Source: §V · p. 5
- One DLS step reduces mean position error by a factor of 4 to below 1% and brings mean orientation error below 1°.After three steps, median error is generally in the 0.01% range, though the 99th-percentile error remains high; 6-DoF robots retain 1.98% position error at that percentile after five steps.Source: §V · p. 5
- One DLS step brings real-world hand-crafted robots to about half a centimeter mean error, and three steps reach sub-millimeter precision.Table II reports raw MorphIK results and results after 1, 3, and 5 DLS steps, using the best of 16 samples.Source: §V · p. 5, Table II · p. 6
- After one DLS step, 8- and 9-DoF wrist variants optimize faster than their corresponding 6- and 7-DoF arms, although wrists have little effect on raw error.With unit-length scaling, position errors range from 2.37% for the Piper with Shadow Hand to 5.98% for the Comau e.DO with Shadowhand wrist; orientation errors range from 1.69° to 4.78°.Source: §V · p. 6
A. Null space Modeling
MorphIK produces diverse in-spec configurations for the same target pose, with the number of distinct solutions varying across robot morphologies and degrees of freedom. The reported spread depends on both the robot’s usable motion and how solutions are counted.
A. Null space Modeling
- The spread evaluation samples 128 solutions per pose, optimizes them in 3 DLS steps, and reports diversity measures only for in-spec solutions.The in-spec criteria are 1% and 2°; distinct0.25 counts solutions differing by at least 0.25 radians in one joint, and std reports their standard deviation.Source: A. Null space Modeling · p. 6, Table III · p. 7
- For synthetic 6-DoF robots, the mean is 3.4 different configurations per pose with a standard deviation of 0.28 rad; distinct counts and standard deviation rise with DoF.6-DoF chains are a special case: they have no null space, but may have up to 16 distinct redundant configurations for a pose.Source: A. Null space Modeling · pp. 6–7
- Real 7-DoF arms yield 42.8 distinct configurations on average, while adding a wrist to a 7-DoF arm yields 80.5 distinct configurations among 96.1 in-spec solutions.For 6-DoF arms, the average is 6.2 distinct configurations; when nearly every sampled solution is distinct, the count is bounded by the number of samples.Source: A. Null space Modeling · p. 7, Table III · p. 7
- 8-DoF arms behave slightly worse than 7-DoF arms because their added wrist has smaller joint ranges and limits motion.The reported 80.5 distinct configurations for a 7-DoF arm with an added wrist are out of 96.1 in-spec solutions.Source: A. Null space Modeling · p. 7
B. Finetuning
The paper evaluates fine-tuning MorphIK for individual real robots against training from scratch. Fine-tuning improves performance quickly, while the from-scratch control narrows the gap with longer training but does not catch up within the tested time frame.
B. Finetuning
- After 50 steps, fine-tuning reduces position error by 1.6% of reach and orientation error by 1.1°; from-scratch training learns nothing in that interval.The reported reductions correspond to about 30% improvement in position and 25% in orientation; 50 steps take 1 second on the tested hardware.Source: B. Finetuning · p. 7, Table IV · p. 7
- After 1,000 steps, fine-tuned position error is halved relative to the baseline, while the from-scratch model still fails on orientation and learns position slowly.The gap between pretrained and from-scratch controls shrinks with longer training.Source: B. Finetuning · p. 7
- After 20,000 steps, mean positional error is 1.48% of reach for the fine-tuned model versus 4.41% for the from-scratch control.The paper reports that the gap has narrowed to a factor of 3, but the control has not caught up within the tested time frame.Source: B. Finetuning · pp. 7–8, Table IV · p. 7
C. Time
Batched inference reduces MorphIK’s per-pose runtime, and the conclusion reports that the model generalizes from procedurally generated training robots to unseen real arms. DLS polishing further reduces residual error, while parallel null-space exploration is a stated advantage.
C. Time
- Batches of 512 poses cost about 0.16 ms per pose, compared with about 3.5 ms for a single pose using best-of-16 inference.DLS adds 5 ms upfront and about 1.7 ms per step; best-configuration selection adds 1.3 ms.Source: C. Time · p. 8
VI. CONCLUSION
- Trained only on procedurally generated arms, MorphIK generalizes to unseen real 6- to 9-DoF arms with about 4% of reach mean positional error and 3.2° orientation error without further setup.After 3 DLS steps, residual error falls below 0.1% of reach and into the 1° range, corresponding to sub-1 mm position error on regular-sized arms.Source: §VI · p. 8
- The conclusion identifies parallel exploration of manipulators’ null spaces as an advantage over traditional non-neural generative methods.Future work will examine the learned embedding space for different tasks and explore a transformer-based decoder for DoF-independent modeling.Source: §VI · p. 8