LiPo: A Lightweight Post-optimization Framework for Smoothing Action Chunks Generated by Learned Policies

Dongwoo Son; Suhan Park

REVIEW 4 major objections 4 minor 20 references

LiPo: A Lightweight Post-optimization Framework for Smoothing Action Chunks Generated by Learned Policies

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that chunk-boundary jerkiness in learned robot policies can be removed by a lightweight post-optimization wrapper, without retraining, and that doing so raises real-world throwing success from 80% to 90%.

desk verdict Sensible post-processing idea, but the experimental evidence is too thin to support the headline improvement claims. read the letter →

arxiv 2506.05165 v1 pith:76JVPBXR submitted 2025-06-05 cs.RO

Dongwoo Son , Suhan Park This is my paper

classification cs.RO

keywords actionchunkingimitationlearningtrajectorysmoothingminimumjerkquadraticprogrammingrobotmanipulationdynamicthrowingquinticsplineinterpolation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the jerkiness of chunked robot actions—sharp discontinuities where one predicted action block ends and the next begins—can be largely removed by a fast post-processing step applied after the learned policy, without retraining. The proposed LiPo framework freezes the motion during inference delay, linearly blends overlapping chunks, then solves a bounded jerk-minimizing quadratic program to refine the trajectory. On a real position-controlled arm throwing a ball into a basket, this raised task success from 80% (raw chunks with quintic spline) to 90%, while visibly reducing position, velocity, and acceleration discontinuities. The payoff would be that dynamic tasks such as throwing, which many imitation-learning policies handle poorly because of stop-and-go chunk execution, can be made usable with a lightweight wrapper instead of a new policy architecture.

What carries the argument

The central object is the bounded-perturbation minimum-jerk quadratic program applied to a linearly blended reference trajectory. Action chunks are $k_a$-step action sequences predicted by a learned policy; LiPo treats each new chunk as a candidate continuation, holds the previous chunk during the inference delay, blends it with the next using a linear weight $\alpha(t)=(t-t_d)/(t_b-t_d)$, and then minimizes $\int_0^T \|d^3(q_{\text{ref}}+\epsilon)/dt^3\|^2 dt$ inside an $\ell^\infty$ box. The box constraints are the load-bearing design choice: they make the optimization a lightweight box-constrained QP, solvable at 30 Hz with an interior-point solver, and they define how much the smoothed path may deviate from the policy's intent. Quintic spline interpolation then raises the low-rate optimized samples to the 400 Hz position-control loop with continuity through the second derivative.

What would settle it

Run the ball-toss comparison with the perturbation bounds set to zero while keeping the blending, delay scheduling, and quintic spline; if the success rate does not drop back toward the raw-chunk baseline, the bounded jerk-minimizing step is not the cause of the improvement. Conversely, a sweep that varies $\bar{\epsilon}_b$ from 0 to 0.1 rad that shows success collapsing at moderate values would falsify the claim that the method is robust.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that smoothing is a post-optimization problem, not a policy-learning problem: chunk-boundary discontinuities can be treated as bounded perturbations of an otherwise good learned trajectory. LiPo constructs a reference by holding the previous chunk during the inference-delay window and linearly blending into the next chunk, then solves the quadratic program that minimizes the integral of the squared third derivative (jerk) of $q_{\text{ref}}(t)+\epsilon(t)$, subject to box constraints $\|\epsilon(t)\|_\infty \le \bar{\epsilon}_b$ in the blending zone and $\|\epsilon(t)\|_\infty \le \bar{\epsilon}_p$ along the rest of the path, with $\epsilon(t)=0$ during the delay. A worst-case task-space bound $\delta x_{\max}(t)=\|J(q(t))\|\,\bar{\epsilon}$ is derived from the Jacobian, so the joint-space perturbation comes with a conservative Cartesian safety estimate. The experiments report that this post-optimized trajectory, executed with quintic spline interpolation, achieves 90% success in the ball-toss task versus 80% for raw chunks with the same spline, and that linear-only interpolation after optimization is worse (60%), which the paper attributes to insufficient smoothness for momentum transfer.

Load-bearing premise

The hand-picked perturbation bounds ($\bar{\epsilon}_b=0.02$ rad in the blending zone, $\bar{\epsilon}_p=0.003$ rad along the path) are assumed to be wide enough to remove discontinuities but small enough to preserve the learned throw, and the paper gives no sensitivity analysis showing this balance holds.

Editorial extensions

If this is right

If the approach generalizes as claimed, any action-chunking policy can be made smoother by wrapping it with LiPo, with no retraining and no change to the policy architecture.
Dynamic tasks like throwing and lifting become more viable for position-controlled, low-cost arms, since the method targets exactly the inference-delay pauses and boundary jumps that break momentum transfer.
The success-rate gap between quintic-spline and linear interpolation after optimization (90% versus 60%) indicates that the smoothing pass must be paired with a high-order interpolation to deliver dynamic performance.
The derived Jacobian-based task-space bound gives a practical runtime safety estimate: the worst-case Cartesian deviation caused by joint-space smoothing is computable from the robot's current configuration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Extension not in the paper: a systematic sweep of $\bar{\epsilon}_b$ and $\bar{\epsilon}_p$ would turn the hand-tuned bounds into a tunable smoothing dial and would reveal whether the 90% success rate is a robust plateau or a narrow peak.
Extension not in the paper: one could test the method's claim about preserving task intent by applying LiPo to a timing-critical task such as flipping an object, where over-smoothing that delays momentum buildup should measurably hurt release accuracy.
Extension not in the paper: the same bounded-perturbation idea could be formulated directly in Cartesian space rather than joint space, exchanging the Jacobian bound for a direct end-effector bound and gaining a more interpretable safety guarantee at some computational cost.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick read for you. The paper proposes LiPo, a post-processing layer that smooths the action chunks emitted by learned policies. The core idea is sensible: instead of executing chunked actions as-is, you overlap consecutive chunks, blend them linearly in the overlap, then run a small bounded-jerk QP to reduce discontinuities while keeping the trajectory close to the policy output. That specific combination—especially the inference-aware overlap scheduling that avoids the stop-and-go caused by policy latency—is not in the cited literature, so as a new application of standard tools it is legitimate.

The paper does several things well. It explains why temporal ensemble is not a good fit (per-timestep inference cost, mode-blending risk). It gives a clear reason for excluding dynamics constraints: they would change the timing the policy assumed. And it includes a Jacobian-based task-space error bound that gives a conservative safety estimate. The writing is clear, and the qualitative plots show visibly smoother position, velocity, and acceleration profiles.

The soft spot is the experimental evidence. Table I is the only quantitative success-rate data: 10 trials per condition, no significance testing, no error bars. The headline comparison, LiPo+quintic at 90% vs raw+quintic at 80%, is statistically indistinguishable from chance (exact two-sided Fisher p=1.0). More telling, LiPo+linear (60%) is worse than raw+linear (70%), so the apparent benefit is attributable to the quintic spline, not to LiPo itself. The paper also provides no numerical smoothness metrics, no sensitivity analysis for the hand-picked bounds (epsilon_b=0.02 rad, epsilon_p=0.003 rad) or time parameters, no repeated policy training seeds, and no code or data. The qualitative single-trajectory figures are suggestive, not demonstrative.

These are real weaknesses, but not fatal to the idea. The method is plausible and cheap, and the bounded-perturbation formulation is reasonable. It just isn't backed by the evidence as written. A practitioner looking for a lightweight smoothing trick could try it; a researcher should treat the results as anecdotal.

My recommendation: send it to peer review—the contribution is meaningful and the presentation is competent—but flag that the experimental section needs major work before the claims can be accepted. More trials, statistical tests, sensitivity analysis, quantitative smoothness metrics, and a comparison that isolates the effect of LiPo from the spline would be needed.

Referee Report

4 major / 4 minor

Summary. The paper proposes LiPo, a lightweight post-optimization layer for action chunks produced by learned policies. The method constructs a linearly blended reference trajectory between a previously optimized chunk and a newly inferred chunk while accounting for inference delay, then solves a jerk-minimizing quadratic program in a bounded perturbation space. Validation is performed on a position-controlled OpenManipulator-Y for a pouch throw and a ball toss; Table I reports success rates from 10 trials per condition, and Fig. 7 shows single-trajectory position, velocity, and acceleration plots with and without LiPo.

Significance. If the claimed effect is real, LiPo is a simple, computationally light, and potentially general post-processing step for chunked action policies. The optimization formulation in Sec. III-B is self-contained and clearly posed as a box-constrained QP, which is a practical strength for real-time control. However, the paper's contribution is primarily empirical, so the quality of the experimental evidence is decisive. The paper also explicitly acknowledges that no robot dynamics constraints are imposed (Sec. V), which is an honest limitation rather than a hidden one.

major comments (4)

[Sec. IV-A and Eq. (4)] The numerical values of td and tb appear to be swapped, which makes the defined blending interval empty. The text sets T=1.66 s, td=0.33 s, and tb=0.16 s, while also saying that these correspond to a 50-step chunk, a 5-step delay, and a 10-step blending horizon. Since 50 steps at 1.66 s give a 30 Hz step, 5 steps correspond to about 0.17 s and 10 steps to about 0.33 s, so td should be about 0.16 s and tb about 0.33 s. As written, the interval (td, tb] in Eq. (4) is empty and the blending region is undefined; this is an internal inconsistency in the central algorithm description and must be fixed.
[Sec. IV-B, Table I] The headline success-rate improvement is not statistically supported. Only 10 trials are reported per condition, with no error bars, no repeated policy training seeds, and no significance test. For the LiPo-quintic versus raw-quintic comparison, 9/10 versus 8/10 successes give a two-sided Fisher exact p-value of 1.0, so the claimed 90% versus 80% advantage is indistinguishable from chance. Moreover, the table shows LiPo with linear spline (60%) performing worse than raw actions with linear spline (70%), so LiPo does not consistently improve success rates. The authors need more trials, confidence intervals, and a defined statistical comparison, or they should soften the success-rate claim.
[Sec. IV-B and Fig. 7] The central qualitative claim that LiPo 'significantly reduces vibration and motion jitter' is not backed by any quantitative smoothness metric. Fig. 7 shows a single representative trial, and no aggregate jerk, acceleration discontinuity, or tracking-error statistics are reported. The paper should report metrics such as RMS jerk, maximum jerk, or acceleration discontinuity count across trials and conditions, so that the smoothness benefit can be evaluated independently of the underpowered success-rate data.
[Sec. III-B and Sec. IV-A] No sensitivity analysis is provided for the hand-picked perturbation bounds eps_b=0.02 rad and eps_p=0.003 rad, or for the timing parameters td and tb. These values control the balance between smoothing and preserving the policy's task-relevant dynamics, so the lack of any sweep leaves the method's robustness unsubstantiated. A sensitivity study over these parameters, even on a subset of trials, is necessary to support the claim that the method preserves task intent while smoothing.

minor comments (4)

[Sec. III-D] The norm used for the Jacobian in Eqs. (8) and (9) is unspecified; the authors should state whether they use the spectral norm, the Frobenius norm, or another matrix norm, and explain how the maximum value is estimated in practice.
[Sec. IV-A] The quintic spline interpolation is described only verbally. The knot spacing, boundary conditions, and how the 30 Hz action chunk is converted to the 400 Hz control signal should be stated precisely for reproducibility.
[Sec. VI] The conclusion section contains two nearly identical 'Future work' paragraphs that should be merged into one.
[Sec. IV-B] The statement that the temporal ensemble (TE) could not be used because the robot 'became unstable' is reported without details; a short description of the observed failure mode would help the reader assess the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the task-success results are external empirical measurements, the optimization constants are hand-set rather than fitted to outcomes, and the paper contains no self-citations.

full rationale

The derivation chain is self-contained against external benchmarks. The reference trajectory qref(t) is built from the learned policy's own overlapping action chunks via Eq. (4)-(5), and the post-optimization (Eq. 6) minimizes the jerk of qref(t)+epsilon(t) within box constraints (epsilon_b=0.02 rad, epsilon_p=0.003 rad) with hand-set timing constants (T=1.66 s, td=0.33 s, tb=0.16 s). None of these parameters is fitted to, or learned from, the reported outcomes, so the headline results in Table I (90% vs 80% success for LiPo+quintic vs raw+quintic, and 60% vs 70% for the linear-spline conditions) are independent physical measurements of task success (ball landing in a basket), not quantities forced by the optimization construction. The paper also contains no self-citations: references [1]-[20] are all external (ACT, Diffusion Policy, UMI, CHOMP/STOMP, minimum-jerk trajectory literature), and no uniqueness theorem or prior-author result is invoked to forbid alternative smoothers. The observed jerk/vibration reduction in Fig. 7 is an expected consequence of minimizing a jerk objective by construction, but the paper does not present that reduction as an empirical discovery separate from its objective; its central empirical claims are the task-success rates and real-trajectory comparisons against raw action chunks. Weaknesses such as the 10-trial sample size, the absence of statistical tests, and the unsupported choice of perturbation bounds are correctness and robustness concerns, not circularity, and per the review rules they are not counted here.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several hand-set parameters (perturbation bounds, blending horizon, inference delay) and on domain assumptions about position control and the effect of jerk minimization. No sensitivity analysis is provided for these choices, and no code or data is released to independently check the experimental results.

free parameters (4)

epsilon_b = 0.02 rad
Perturbation bound in the blending zone, chosen by hand; no sensitivity analysis is provided (Section III.B, Section IV.A).
epsilon_p = 0.003 rad
Perturbation bound in the path region, chosen by hand; no sensitivity analysis (Section III.B, Section IV.A).
tb = 0.16 s (10 steps)
Blending horizon length, set to 10 steps; no justification for this value (Section IV.A).
td = 0.33 s (5 steps)
Inference delay window, set to 5 steps; assumed constant (Section IV.A).

assumptions (5)

domain assumption The policy outputs joint position commands, and the robot is position-controlled, so smoothing positions directly affects executed motion.
Section IV.A states the manipulator is used in position control mode.
domain assumption Minimizing jerk improves smoothness and reduces mechanical wear.
Stated in Introduction and Related Work as motivation for jerk minimization.
standard math A first-order Taylor approximation of forward kinematics provides a conservative bound on task-space deviation.
Equation (8) uses the Jacobian norm bound.
domain assumption The inference delay td is a known constant (0.33 s) during execution.
Section IV.A sets td to 5 steps and treats it as fixed, but real inference delays may vary.
standard math The quadratic program with box constraints is convex and can be solved quickly by Clarabel.
The problem formulation in Section III.B is a box-constrained QP, assumed solvable in real time.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LiPo: A Lightweight Post-optimization Framework for Smoothing Action Chunks Generated by Learned Policies." pith.science (2026). https://pith.science/paper/76JVPBXR

@misc{pith2026250605165,
  author       = {Pith},
  title        = {Pith review of: LiPo: A Lightweight Post-optimization Framework for Smoothing Action Chunks Generated by Learned Policies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/76JVPBXR}},
  note         = {Machine review of arXiv:2506.05165}
}

read the original abstract

Recent advances in imitation learning have enabled robots to perform increasingly complex manipulation tasks in unstructured environments. However, most learned policies rely on discrete action chunking, which introduces discontinuities at chunk boundaries. These discontinuities degrade motion quality and are particularly problematic in dynamic tasks such as throwing or lifting heavy objects, where smooth trajectories are critical for momentum transfer and system stability. In this work, we present a lightweight post-optimization framework for smoothing chunked action sequences. Our method combines three key components: (1) inference-aware chunk scheduling to proactively generate overlapping chunks and avoid pauses from inference delays; (2) linear blending in the overlap region to reduce abrupt transitions; and (3) jerk-minimizing trajectory optimization constrained within a bounded perturbation space. The proposed method was validated on a position-controlled robotic arm performing dynamic manipulation tasks. Experimental results demonstrate that our approach significantly reduces vibration and motion jitter, leading to smoother execution and improved mechanical robustness.

Figures

Figures reproduced from arXiv: 2506.05165 by the authors.

**Figure 3.** End-effector setup used in the experiments. The [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Camera input preprocessing: (a) original image and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 6.** Figure 6: Visualization of action chunk smoothing with LiPo. Colored segments represent different action chunks predicted [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: Comparison of joint position, velocity, and acceleration trajectories with and without the proposed LiPo method. [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 13 canonical work pages

[1]

An algorithmic perspective on imitation learning,

T. Osa, J. Pajarinen, G. Neumann, J. A. Bagnell, P. Abbeel, J. Peters et al., “An algorithmic perspective on imitation learning,” Foundations and Trends® in Robotics , vol. 7, no. 1-2, pp. 1–179, 2018

work page 2018
[2]

Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,

X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne, “Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,” ACM Transactions on Graphics (TOG) , vol. 37, no. 4, pp. 1–14, 2018

work page 2018
[3]

Learning fine-grained bimanual manipulation with low-cost hardware,

T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” in Robotics: Science and Systems (RSS) , 2023

2023
[4]

Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,

Z. Fu, T. Z. Zhao, and C. Finn, “Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,” in Conference on Robot Learning (CoRL) , 2024

2024
[5]

Bi-act: Bilateral control-based imitation learning via action chunking with transformer,

T. Buamanee, M. Kobayashi, Y . Uranishi, and H. Takemura, “Bi-act: Bilateral control-based imitation learning via action chunking with transformer,” in 2024 IEEE International Conference on Advanced Intelligent Mechatronics (AIM) . IEEE, 2024, pp. 410–415

work page 2024
[6]

InterACT: Inter-dependency aware action chunking with hierarchical attention transformers for bimanual manipulation,

A. C.-W. Lee, I. Chuang, L.-Y . Chen, and I. Soltani, “InterACT: Inter-dependency aware action chunking with hierarchical attention transformers for bimanual manipulation,” in 8th Annual Conference on Robot Learning , 2024. [Online]. Available: https://openreview.net/ forum?id=lKGRPJFPCM

work page 2024
[7]

Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,

C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song, “Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,” in Proceedings of Robotics: Science and Systems (RSS) , 2024

work page 2024
[8]

Diffusion policy: Visuomotor policy learning via action diffusion,

C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” The International Journal of Robotics Research , 2024

2024

Show all 20 references

[9]

UMI on legs: Making manipulation policies mobile with manipulation-centric whole-body controllers,

H. Ha, Y . Gao, Z. Fu, J. Tan, and S. Song, “UMI on legs: Making manipulation policies mobile with manipulation-centric whole-body controllers,” in Proceedings of the 2024 Conference on Robot Learn- ing, 2024

2024
[10]

Stomp: Stochastic trajectory optimization for motion planning,

M. Kalakrishnan, S. Chitta, E. Theodorou, P. Pastor, and S. Schaal, “Stomp: Stochastic trajectory optimization for motion planning,” in 2011 IEEE international conference on robotics and automation . IEEE, 2011, pp. 4569–4574

2011
[11]

Chomp: Covariant hamiltonian optimization for motion planning,

M. Zucker, N. Ratliff, A. D. Dragan, M. Pivtoraiko, M. Klingensmith, C. M. Dellin, J. A. Bagnell, and S. S. Srinivasa, “Chomp: Covariant hamiltonian optimization for motion planning,” The International journal of robotics research , vol. 32, no. 9-10, pp. 1164–1193, 2013

2013
[12]

On learning, representing, and generalizing a task in a humanoid robot,

S. Calinon, F. Guenter, and A. Billard, “On learning, representing, and generalizing a task in a humanoid robot,” IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) , vol. 37, no. 2, pp. 286–298, 2007

2007
[13]

Minimum jerk for trajectory planning and control,

K. J. Kyriakopoulos and G. N. Saridis, “Minimum jerk for trajectory planning and control,” Robotica, vol. 12, no. 2, pp. 109–113, 1994

1994
[14]

Global minimum-jerk trajectory planning of robot manipulators,

A. Piazzi and A. Visioli, “Global minimum-jerk trajectory planning of robot manipulators,” IEEE Transactions on Industrial Electronics , vol. 47, no. 1, pp. 140–149, 2000

2000
[15]

Global minimum-jerk trajec- tory planning of space manipulator,

H. Panfeng, X. Yangsheng, and L. Bin, “Global minimum-jerk trajec- tory planning of space manipulator,” International Journal of Control, Automation, and Systems , vol. 4, no. 4, pp. 405–413, 2006

2006
[16]

Smooth trajectory planning for a parallel manipulator with joint friction and jerk constraints,

L. Liu, C. Chen, X. Zhao, and Y . Li, “Smooth trajectory planning for a parallel manipulator with joint friction and jerk constraints,” International Journal of Control, Automation and Systems , vol. 14, no. 4, pp. 1022–1036, 2016

2016
[17]

On the performance of jerk-constrained time-optimal trajectory planning for industrial manipulators,

J.-e. Lee, A. Bylard, R. Sun, and L. Sentis, “On the performance of jerk-constrained time-optimal trajectory planning for industrial manipulators,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 9772–9778

2024
[18]

ROBOTIS OpenManipulator-Y,

ROBOTIS Co., Ltd., “ROBOTIS OpenManipulator-Y,” https://www. dynamixel.com/omy.php, accessed: 2025-05-25

2025
[19]

Clarabel: An interior-point solver for conic programs with quadratic objectives,

P. J. Goulart and Y . Chen, “Clarabel: An interior-point solver for conic programs with quadratic objectives,” 2024

2024
[20]

Lerobot: State-of-the-art machine learning for real-world robotics in pytorch,

R. Cadene, S. Alibert, A. Soare, Q. Gallouedec, A. Zouitine, and T. Wolf, “Lerobot: State-of-the-art machine learning for real-world robotics in pytorch,” https://github.com/huggingface/lerobot, 2024

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

An algorithmic perspective on imitation learning,

T. Osa, J. Pajarinen, G. Neumann, J. A. Bagnell, P. Abbeel, J. Peters et al., “An algorithmic perspective on imitation learning,” Foundations and Trends® in Robotics , vol. 7, no. 1-2, pp. 1–179, 2018

work page 2018

[2] [2]

Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,

X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne, “Deepmimic: Example-guided deep reinforcement learning of physics-based char- acter skills,” ACM Transactions on Graphics (TOG) , vol. 37, no. 4, pp. 1–14, 2018

work page 2018

[3] [3]

Learning fine-grained bimanual manipulation with low-cost hardware,

T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” in Robotics: Science and Systems (RSS) , 2023

2023

[4] [4]

Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,

Z. Fu, T. Z. Zhao, and C. Finn, “Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,” in Conference on Robot Learning (CoRL) , 2024

2024

[5] [5]

Bi-act: Bilateral control-based imitation learning via action chunking with transformer,

T. Buamanee, M. Kobayashi, Y . Uranishi, and H. Takemura, “Bi-act: Bilateral control-based imitation learning via action chunking with transformer,” in 2024 IEEE International Conference on Advanced Intelligent Mechatronics (AIM) . IEEE, 2024, pp. 410–415

work page 2024

[6] [6]

InterACT: Inter-dependency aware action chunking with hierarchical attention transformers for bimanual manipulation,

A. C.-W. Lee, I. Chuang, L.-Y . Chen, and I. Soltani, “InterACT: Inter-dependency aware action chunking with hierarchical attention transformers for bimanual manipulation,” in 8th Annual Conference on Robot Learning , 2024. [Online]. Available: https://openreview.net/ forum?id=lKGRPJFPCM

work page 2024

[7] [7]

Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,

C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song, “Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,” in Proceedings of Robotics: Science and Systems (RSS) , 2024

work page 2024

[8] [8]

Diffusion policy: Visuomotor policy learning via action diffusion,

C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” The International Journal of Robotics Research , 2024

2024