acceptodds
Under review as a conference paper at ICLR 2027

Kine2Go: Kinematic dataset for the Unitree Go2 robot with diverse gaits and motions

Abstract

The recent popularity of robotics, combined with the steadily decreasing cost of robotic hardware, has lowered the entry barrier to robotics research and enabled rapid advancements in the field. One of the primary examples is the Unitree Go2 quadruped robot, which is often used by researchers in the areas of locomotion, navigation, control, and others. Many researchers use the Go2 robot in combination with techniques like imitation learning, reinforcement learning, and behavioral cloning to allow machine learning systems to take full control of the robot. At the same time, many of those techniques require demonstration data consisting of the robot's kinematics information and actions applied to the motors. Obtaining such data is difficult, requires building complex pipelines, and can take significant time. To aid in those kinds of efforts, we present Kine2Go, a state-action locomotion dataset for the Unitree Go2 robot comprising 40 motion families, each associated with an RL tracking policy and represented by 20 stochastic rollouts. Our pipeline accepts data from various quadruped morphologies and translates them to a Go2-compatible format. Then we use Reinforcement Learning to train policies following a given motion, and finally we gather data from those policies, which yields perturbed kinematic trajectories with corresponding motor-level actions. To showcase the value of our dataset, we also train a Behavioral Foundation Model (BFM) on it and evaluate its motion-tracking and zero-shot reward-conditioning capabilities.

Primary area:
datasets and benchmarks
Keywords:
Robotics, Dataset, Reinforcement Learning, Motion Imitation, Kinematics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bCKeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Kine2Go, a simulated state–action dataset for low-level quadrupedal locomotion. The proposed pipeline retargets heterogeneous quadruped motions to the Unitree Go2, trains RL tracking policies, and collects trajectories with corresponding motor commands. The dataset contains 40 motion families and 800 rollouts. Its utility is demonstrated through motion tracking and zero-shot reward-conditioned control with a Behavioral Foundation Model (BFM).

Critical strengths
  • Providing motor actions alongside kinematic trajectories makes the dataset useful for learning control policies, beyond using motion references alone.
  • The pipeline brings together several motion sources in a common Go2 representation. The BFM experiments and motion-family ablation provide an initial demonstration of its utility.
Critical weaknesses
  • The released state–action trajectories are generated entirely in simulation on flat terrain, and neither the tracking policies nor the BFM are evaluated on hardware. Although this limitation is acknowledged, the dataset’s practical value for real-world locomotion remains unverified.
  • The contribution is more convincing as a reusable data-generation pipeline than as a broadly applicable dataset. The dataset targets a single robot morphology, and its 800 trajectories are repeated executions of 40 motion families. It is unclear whether this provides sufficient coverage for the broader locomotion-learning applications discussed. Extending the pipeline to another target robot would also require retargeting and controller training; the claim that this needs only “minor modifications” is not demonstrated.
  • The BFM experiments establish one use case, but provide limited evidence of the dataset’s broader value. The motion-family ablation is useful, yet motion tracking and reward-conditioned control within the same simulation setup do not establish how well the data supports learning beyond the collected behaviors. The broader claims about enabling general locomotion policies should be qualified accordingly.
Questions for authors
  • What changes and retraining would be needed to use the Go2 motions to construct a dataset for another quadruped? Can the authors clarify which components are reusable and which are specific to Go2?
  • How much additional state–action coverage do the 20 rollouts per family provide after accounting for randomized heading? Section 5 distinguishes execution diversity from behavioral diversity, but the planar visualization alone does not quantify this. Could the authors provide a coverage or redundancy analysis?
  • Beyond the BFM demonstration, what evidence would establish the value of the released state–action pairs for the other stated uses, such as behavioral cloning or offline control?
1 comment

We thank the reviewer for the thoughtful review and concrete questions. Numbers refer to the revised manuscript.

Real-world validation

We agree that the dataset's real-world value is not verified. We now state explicitly that all experiments are in simulation; hardware validation was not feasible within the discussion period.

Pipeline vs. dataset; other morphologies (Question 1)

We agree that the claim about "minor modifications" was not supported. We removed it and added Appendix J with Table 14. It describes which components are reusable and which are specific to the Go2:

  • Reusable: the source adapters (common body, hip and foot keypoints); trunk-pose estimation and multi-link foot IK; the PPO setup with future-reference conditioning, motion cycling and tracking rewards; rollout collection and review tools.
  • Go2-specific: the robot description and the trunk, hip, foot and joint correspondence; the calibration of motion scale, trunk height and foot offsets; fixed assumptions about 12 joints, 4 feet, joint order and observation and action dimensions; actuator, PD, termination and reward-tolerance settings.
  • Must be regenerated: all references, expert policies and rollouts.

For a quadruped with a rigid trunk and three actuated joints per leg, we estimate a moderate engineering effort. A different joint count or topology would require more substantial changes. We state explicitly that we have not ported the pipeline to another robot.

Coverage of the 20 rollouts per family (Question 2)

Appendix G quantifies within-family execution diversity. It removes each trajectory's initial position and heading before computing pairwise differences in joints, actions, feet, root position and heading, so the reported diversity is not an artifact of the random initial heading. To measure what these executions add for learning, we use the new behavioral cloning experiment (Appendix I, Table 13). We train reference-conditioned BC on one family from each source, using either 1 or all 20 rollouts, with the same update budget. With 1 rollout, BC fails within about 0.8 s in nearly all episodes (success 0.012). With 20 rollouts, survival rises to 4.0 s and success to 0.313. Since the policy observation contains the absolute root pose, part of this gain could come from heading coverage alone. We therefore attempted a control that rotates and translates the single rollout to the same 20 initial placements as the 20-rollout data, which gives the same heading coverage and amount of data but only one execution. With this augmentation, BC reaches an average survival of 2.1 s and a success fraction of 0.137, between the 1-rollout (0.8 s, 0.012) and the 20-rollout policies (4.0 s, 0.313). For the circle walk, augmentation gives almost no improvement over a single rollout (0.78 vs. 0.67 s, compared with 4.26 s for 20 rollouts). However, the transformed observations are not fully consistent with the expert: the expert's actions on them differ from the recorded labels by 0.2 to 1.9 normalized units on average, depending on the family. This suggests that the absolute-pose policies are not exactly invariant to re-placement, so the augmented labels are partly inconsistent. We therefore consider this control preliminary and have not included it in the paper. It suggests that heading coverage explains part, but not all, of the gain from additional rollouts.

Value beyond the BFM use case (Question 3)

The revision adds three pieces of evidence for the usefulness of our work beyond the original BFM experiments:

  1. Recorded actions: in the behavioral cloning proof of concept above, the actions can be learned directly, and more executions help substantially. Plain BC without augmentation remains well below the experts (0.971), and we present it as a proof of concept, not a strong baseline. We have not evaluated offline RL.
  2. Generalization beyond the training motions (Appendix F.4): on families held out from training, the 10 and 20 family models reach success fractions of 0.563 and 0.596, compared with 0.758 for the 40 family model trained on them. Generalization is good for held-out clips with a close relative in training and poor for distinct gaits and Solo8.
  3. Value of physically (in simulation) executed data (Appendix F.5): a BFM regularized on the rollouts tracks the kinematic references about as well as one regularized on the references themselves (success 0.743 vs. 0.726). The reverse does not hold (0.623 vs. 0.791 on the rollouts).

We also qualified the broader claims: we added a contribution list, replaced the AMASS/MoCapAct comparison and the "comprehensive" wording, qualified the statements on scalability, hardware and general locomotion policies, and added a limitation on behavioral coverage and within-source generalization.

Reviewer_taBjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Kine2Go which is a dataset for the Unitree Go2. Motions from four different sources are retargeted to the Go2 with frame-wise IK. The authors train for every motion a PPO tracking policy in Genesis and then collect 20 rollouts per policy which gives 40 motion families and 800 trajectories. As a use case, the authors adapt Meta Motivo to the Go2, train it with Kine2Go and report EMD and tracking metrics. There are no real robot experiments in the paper.

Critical strengths

The BFM experiment is a reasonable and interesting proof of concept.

Critical weaknesses

The dataset is too small and not diverse enough, and the contribution besides the dataset is really limited. 800 trajectories is very little especially as the 20 rollouts of a family come from the same policy and mainly differ by the random heading. So the real diversity is 40 clips, and many of them are very short loops which are just repeated over the whole rollout. Also 22 of 40 families come from literally the same dog source as both the wolf demo and the dog clips from Peng et al use the same mocap data. In my opinion this is far away from being "analogous to AMASS or MoCapAct". The pipeline itself is a re-implementation of Peng et al. / MimicKit in Genesis, which is not a contribution at all. Compared to all the other existing Go2 retargeting works and datasets, this one adds no novelty and no reasonable contribution. The presentation of the work and writing also needs heavy improvement. Also no real world experiments is not acceptable for a quadruped locomotion paper like this.

Questions for authors

In the 10/20-family ablation, is the evaluation done on all 40 families or only on the training families?

1 comment

We thank the reviewer for the review. We address each point below. Section, table and appendix numbers refer to the revised manuscript.

Question: which families are the 10 and 20 family models evaluated on?

In Figure 5, all models, including the 10 and 20 family ones, are evaluated against the full dataset (40 families plus the standing trajectory). The reduced-corpus curves therefore mix training and held-out families. We now state this in the caption and in Section 5.1. Appendix F.4 (Tables 7–8) reports training and held-out families separately. On the 20 families held out from both reduced corpora, the 10 and 20 family models reach success fractions of 0.563 and 0.596, compared with 0.758 for the 40 family model trained on them. Held-out families with a closely related family in training, such as the VHDC gaits, are tracked almost as well as by the full model. Held-out Solo8 motions and distinct gaits are not. We report this as generalization within the sources seen in training and list it as a limitation.

Dataset size and diversity

We agree that the dataset is small and that its 800 trajectories are 20 executions of each of 40 motion families. The revised manuscript states this explicitly. It separates behavioral diversity (across families) from execution diversity (within families; quantified in Appendix G) and adds a limitation on behavioral coverage. We also removed the statement that Kine2Go is "analogous to AMASS or MoCapAct". It now says that Kine2Go is much smaller and is intended to play a similar role for the Go2.

The repeated executions are not redundant for learning, however. In our new behavioral cloning experiment (Appendix I), BC trained on one rollout per family fails within about 0.8 s (success 0.012). Trained on all 20 rollouts, it survives 4.0 s on average (success 0.313). Part of this gain may come from the wider coverage of initial headings rather than from execution diversity alone, as discussed in Appendix I and in our response to Reviewer 3. Our new comparison with regularization on the kinematic references alone (Appendix F.5) also shows that the physically executed rollouts are a better regularizer than the references from which they were produced: success on the rollouts is 0.791 vs. 0.623, at similar accuracy on the references.

AI4Animation sources

The reviewer is correct that the 15 AI4Animation mocap families and the 7 families from the interactive demo both derive from the same dog motion-capture data. The interactive demo is still a distinct source of motions: it produces new, user-steered trajectories (circles, figure-eights, squares, ellipses and strafing paths) that do not appear in the original clips. To make the composition clear, all new analyses report results per source (Tables 8, 10 and 11), and these show clear differences between the two AI4Animation subsets.

Novelty of the pipeline

We agree that the retargeting-and-imitation methodology is not new, and Section 3 says so explicitly. Our contribution is the released dataset of Go2 state-action rollouts, together with the expert policies and a Genesis-based pipeline. It is now supported by an evaluation of what the data enables: BFM training with held-out evaluation, a comparison with kinematic references only, a calibration against the experts, and a behavioral cloning proof of concept. The revised introduction lists these contributions explicitly. Section 2.3 discusses the closest prior Go2 work (Liu et al.) and how Kine2Go differs from it: Kine2Go releases state-action rollouts and per-motion policies rather than a generative motion model.

Presentation

We revised the presentation: there is an explicit contribution list, metric definitions and aggregation are clarified (Appendix F.3), and claims are softened throughout (introduction, Sections 3, 5 and 6, limitations).

Real-world experiments

We acknowledge that the paper contains no hardware experiments. The dataset is intended for training in simulation, where BFMs such as Meta Motivo are also trained and evaluated. We now state explicitly that all experiments are in simulation. Hardware validation was not possible within the discussion period, and we list it as an important direction for future work.

Reviewer_YgwtOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Kine2Go, a state-action locomotion dataset for the Unitree Go2 quadruped, and the Genesis-based pipeline used to build it. The pipeline has three stages (Section 3). First, frame-wise IK retargets motions from four sources: AI4Animation dog mocap, VHDC horse mocap, Solo8 robot motions, and a modified AI4Animation wolf animation demo. Second, one PPO tracking policy is trained per reference clip, following Peng et al. (2020), for about 2B environment steps. Third, each policy is rolled out stochastically, and the rollouts are filtered by manual video inspection: 20 rollouts are kept out of 60 (Appendix D). The released dataset has 40 motion families × 20 rollouts = 800 trajectories, most of them 571 frames long at 60 Hz (Table 2). Each frame records joint positions and velocities, base orientation and angular velocity, global link positions and velocities, and normalized PD-target actions.

To show that the data is useful, the authors adapt Meta Motivo (FB-CPR) to the Go2 and train it with Kine2Go as the regularizing dataset. Figure 5 reports tracking EMD during training, together with an ablation over 10, 20 and 40 families, a larger model and a "no data" run. Tables 4–5 report per-family tracking results, and Figure 6 shows zero-shot reward-conditioned control for forward velocity, base height and yaw rate.

The authors claim that the dataset is diverse and fills a gap for quadruped BFMs, that the pipeline runs on consumer hardware and extends easily, and that Kine2Go supports BFM training. The evidence supports a narrower claim: an FB-CPR agent trained with these 800 trajectories can track the same motion families in the Genesis simulator. Generalization, comparative data value and real-robot usefulness are not established.

Critical strengths
  1. The paper targets a real resource gap. MoCapAct and AMASS have no open equivalent for quadrupeds, i.e. a set of retargeted reference motions paired with motor-level actions for a widely used commercial quadruped. The data format records engine-agnostic states and actions together with simulator snapshots for exact replay (Section 3.3). The authors also commit to releasing the code and the dataset (Reproducibility Statement), which could lower the entry cost for quadruped imitation and BFM work.
  2. The pipeline and data provenance are documented in useful detail. Appendix A (Table 1) gives source files and frame ranges per policy. Appendix C (Table 3, Eqs. 1–8) gives the full PPO and reward configuration. Appendix D describes the manual filtering protocol, including the 60-to-20 selection and per-family success rates. These details make the construction process largely reproducible apart from the human selection step, which the authors acknowledge.
  3. The BFM case study includes 5-seed training curves with standard deviations and an ablation over the number of motion families (Figure 5). Per-family results are also reported for both model sizes (Tables 4–5). The "no data" run diverges while the 10, 20 and 40-family runs improve monotonically, which is at least a first indication that the data is used as a meaningful regularizer.
Critical weaknesses

Metric definitions and reporting need tightening. The success fraction uses an unusual threshold of 1.25 on a unitless concatenation of positions and 6D orientations (Appendix E.2). The aggregated link-velocity error is reported in mm/s with values above 6000. Neither metric is calibrated against a reference, e.g. the expert policies' own scores. Tables 4–5 include a "standing" family that does not appear in the 40-family dataset (Tables 1–2). Their standard deviations are computed over the 20 trajectories rather than over training seeds. Figure 6 has no baseline, and for the yaw target of 1.0 rad/s the baseline model reaches about 2.2 rad/s, an error the text does not discuss.

Questions for authors
  1. In Figure 5, which motions are the 10-family and 20-family models evaluated on? Can you report tracking on held-out families, for example leave-one-source-out or held-out clips within a source, so that generalization beyond the training motions is measured?
  2. How does the FB-CPR agent compare when regularized only on the retargeted kinematic references, without the RL rollouts and actions? This comparison is central to the claim that the state-action rollouts, and not just the retargeted references, are the valuable contribution.
  3. For every family, what is the tracking error of the per-motion expert policy against its reference, and what was the acceptance rate in the manual 60-to-20 selection? How consistent were the two annotators?
  4. What are the licenses of the source data (AI4Animation repository data and Unity demo, VHDC, the Solo8 dataset of Li et al. 2023)? Do they permit redistribution of retargeted derivatives, and what license will Kine2Go and its code carry? Will a datasheet or data card be provided?
  5. What PD gains and actuator model were used? Have you checked that the recorded actions reproduce the trajectories in another simulator, such as MuJoCo or Isaac, given the "engine-agnostic" claim?
2 comments

We thank the reviewer for the detailed and constructive review. We address every weakness and question below. Section, table and appendix numbers refer to the revised manuscript.

Metric definitions and reporting

  • Calibration of the success threshold (Appendix H, Tables 11–12). We scored the per-family expert policies against their own reference motions, using the same metric code as in the BFM evaluation. At the threshold of 1.25, the experts reach an average success fraction of 0.956 over 38 families, and 33 of the 38 families exceed 0.95. Table 11 adds a threshold sweep: the expert success fraction is 0.737, 0.875, 0.934, 0.956, 0.966 and 0.974 at thresholds 0.5, 0.75, 1.0, 1.25, 1.5 and 2.0. The threshold is therefore met by accurate expert tracking for almost all families. The exceptions are three fast running motions (canter, run_00, run_02, about 0.5), which are also the hardest families for the BFM. For orientation, the baseline BFM reaches 0.807 success on the same 38 families. Two families with an unrecorded randomized reference start are excluded.
  • Velocity units. Link-velocity error is now averaged per links, and reported in m/s, this and all other metrics are defined in Appendix F.3.
  • "Standing". This is a stationary rollout with zero action commands. It is included in all training corpora, but it is not one of the 40 motion families. It is now explained in Appendix F.2, reported separately in Tables 5–6, and excluded from all averages.
  • Standard deviations. These are now computed across five training seeds, after averaging the 20 trajectories of each family within a seed.
  • Figure 6 and the yaw target. We added a discussion in Appendix F.3. For the yaw rate, the large model stays within about 0.3 rad/s of every target. The baseline model overshoots, reaching about 2.2 rad/s for a target of 1.0 rad/s and −1.33 rad/s for −1.0 rad/s. The small spread across evaluation seeds shows that the error is systematic. We attribute it to an inaccurate reward embedding: the latent is inferred from reward-weighted corpus transitions, and these can be dominated by faster turning behaviors when few transitions match the target. We did not have time to add an external baseline for reward-conditioned control during the discussion period. The large model serves as a within-method comparison.

Q1: Evaluation of the 10 and 20 family models and held-out families

In Figure 5, all models, including the 10 and 20mfamily ones, are evaluated against the full dataset (40 families plus standing), so the reduced-corpus curves mix training and held-out families. This is now stated in the caption and in Section 5.1. Appendix F.4 (Tables 7–8) separates the two. The nested subsets partition the families into G1 (in all corpora), G2 (held out from the 10-family corpus only) and G3 (held out from both reduced corpora). Results:

  • On G3, the 10- and 20-family models reach success fractions of 0.563 and 0.596, compared with 0.758 for the 40 family model trained on them. Excluding the two hardest families, the values are 0.618, 0.654 and 0.827.
  • On G2, the 10-family model reaches 0.606, compared with 0.912 for the 20-family model.
  • Generalization depends on the source. Held-out VHDC families, each of which has a sibling with the same horse and gait in training, are tracked as well as by the full model (0.994 / 0.972 vs. 0.989). Held-out Solo8 families are not tracked as good (0.187 / 0.098 vs. 0.847), and neither are gaits without a close counterpart (0.170 / 0.261 vs. 0.457).

We therefore describe the result as generalization to new clips within the sources seen in training, and we list it as a limitation. We did not run leave-one-source-out training: it requires training new BFMs, which was not feasible within the discussion period.

Q2: Regularization on kinematic references only

We trained the same baseline BFM with the 40 retargeted references as regularization data instead of the 800 rollouts. Architecture, hyperparameters and training steps are identical; we used 3 seeds because of time and compute constraints. Both models are evaluated on both target sets (Appendix F.5, Tables 9–10; the training curve is added to Figure 5).

Regularization dataTargets: rollouts (success / EMD)Targets: references (success / EMD)
Rollouts0.791 / 0.6500.743 / 0.525
References only0.623 / 0.8490.726 / 0.454

The rollout-trained model tracks the references about as well as the reference-only model: higher success and lower relative position error (2.75 vs. 2.91 cm), but higher EMD and velocity error. The reference-only model is much worse on the rollouts. The gap comes mainly from AI4Animation mocap (0.626 vs. 0.413) and Solo8 (0.791 vs. 0.469), and is negligible for VHDC (0.992 vs. 0.976). On reference targets, the rollout-trained model is far better for Solo8 (0.670 vs. 0.437), while the reference-only model is slightly better for the other three sources.

FB-CPR regularizes on states only and does not use dataset actions. This comparison therefore evaluates the physically realized states, not the actions. To test the actions directly, we added a behavioral cloning proof of concept (Appendix I). BC trained on 1 rollout fails within about 0.8 s (success 0.012), while BC trained on 20 rollouts survives 4.0 s on average (success 0.313). Both remain well below the experts (0.971).

Q3: Expert tracking error, acceptance rates and annotators

Per-family expert errors are given in Table 12 (Appendix H); see the first point above. On acceptance rates: as described in Appendix E, we generated 60 rollouts per policy and stopped once 20 acceptable ones were found. The rejected rollouts were not kept, and exact per-family acceptance counts were not recorded. Qualitatively, acceptance was close to 100% for the VHDC families and about 50% for the hardest one (ai4_dog_canter). Each family was reviewed by one of two annotators, so inter-annotator agreement was not measured. We acknowledge this limitation of the manual filtering. The expert-calibration numbers in Appendix H quantify the quality of the retained data with an objective metric.

Q4: Licenses and data card

Appendix A.1 now lists the license of every subset. Each subset inherits the license of its source: CC BY-NC 4.0 for the 22 AI4Animation-derived families, CC BY-SA 4.0 for VHDC (upstream CC BY-SA 3.0, upgraded as its share-alike clause permits), and BSD-3-Clause for Solo8 (the license of the repository that distributes the data; we also obtained the authors' permission). All of these licenses permit redistribution of derived data. The AI4Animation subsets are restricted to non-commercial use, and the VHDC subset is share-alike. The license of every clip is recorded in its metadata. The dataset card (README) lists the license and attribution of each source and includes the full license texts. The documentation is under CC BY 4.0, the helper scripts are under MIT, and the pipeline code is under BSD-3-Clause.

Q5: PD gains, actuator model and cross-simulator replay

Appendix D gives the details. We use explicit PD torque control with Kp = 30 N·m/rad and Kd = 1.5 N·m·s/rad. Policies run at 60 Hz and PD torques are recomputed at 240 Hz. Torque limits are ±23.7 N·m for the hip and thigh joints and ±35.55 N·m for the calf joints. No learned actuator model is used. Appendix D also describes the action scaling, the expert action delay, and the gain randomization used during expert training.

We have not evaluated action replay or policy transfer in another simulator. We revised Section 3.3 to make clear that "simulator-independent" refers only to the numerical data representation, not to equivalent dynamics. For exact replay, each trajectory additionally includes a Genesis simulator snapshot.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all reviewers for their careful reading of our work and constructive feedback. Several concerns were shared across reviews: how well the BFM generalizes beyond its training motions, whether the RL rollouts add value over the retargeted kinematic references, whether the recorded actions are useful beyond BFM training, how the metrics should be interpreted, and whether some claims were stated too broadly. We have uploaded a revised manuscript. The main changes are summarized below; individual responses give relevant details and report important numbers.

New experiments

  • Held-out motion families (Appendix F.4, Tables 7–8). All models in Figure 5, including the 10- and 20-family ones, were evaluated against the full dataset (40 families plus standing). We now report the 10, 20 and 40 family models separately on disjoint groups of training and held-out families. On the 20 families held out from both reduced corpora, the 10 and 20 family models reach success fractions of 0.563 and 0.596, compared with 0.758 for the 40-family model trained on them. Generalization is strong for held-out clips with a close relative in training (e.g., VHDC gaits) and weak for Solo8 and for distinct gaits. We report this as within-source generalization and list it as a limitation.
  • Kinematic references vs. RL rollouts (Appendix F.5, Tables 9–10, Figure 5). We trained the same BFM on the 40 retargeted references instead of the rollouts and cross-evaluated both models on both target sets. The model trained on rollout tracks the references about as well as the reference-only model (success 0.743 vs. 0.726), while the reference-only model is much worse on the rollouts (0.623 vs. 0.791). The gap comes mainly from AI4Animation mocap and Solo8 (which are generally harder clips than VHDC).
  • Expert calibration (Appendix H, Tables 11–12). We scored the per-family expert policies against their own references with the same metric code. At the threshold of 1.25, the experts reach 0.956 success averaged over 38 families. We also report a threshold sweep and per-family values.
  • Behavioral cloning proof of concept (Appendix I, Table 13). We trained reference-conditioned BC on one family from each source, using 1 or 20 rollouts. With 1 rollout, BC fails within about 0.8 s (success 0.012). With 20 rollouts, survival rises to 4.0 s and success to 0.313, still well below the experts (0.971). Part of this gain may come from the wider coverage of initial headings; Appendix I discusses this.

Revised claims and presentation

  • We added an explicit contribution list and state that all experiments are in simulation.
  • We replaced the comparison to AMASS/MoCapAct and the "comprehensive gait collection" wording. We also qualified the claims about scalability, consumer hardware and general locomotion policies, and added a limitation on the dataset's behavioral coverage.
  • Other morphologies (Appendix J, Table 14). We replaced "after minor modifications" with a description, of which components are reusable and which are Go2-specific. We state explicitly that no port to another robot was performed.
  • Metrics and reporting (Appendix F.3, Tables 5–6). Standard deviations are now across five training seeds. Link-velocity errors are in m/s. We added the relative link-position error. The standing trajectory is now explained, and is reported separately and excluded from averages.
  • Additional details. The PD control and actuator model is in Appendix D. Licenses for every data subset and for the code are in Appendix A.1. Within-family execution diversity is quantified in Appendix G. The zero-shot yaw-rate overshoot of the baseline model is discussed in Appendix F.3.

Limitations we could not address in the discussion period

We have no hardware experiments. We also did not run leave-one-source-out BFM training or cross-simulator action replay. We state these limitations explicitly in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.