Pith. sign in

REVIEW 3 major objections 6 minor 18 references

SymmGrid: Super-Scaling On-Robot Learning with Parallelized Symmetries and Egocentric-Exocentric Visual Perception

T0 review · 3 major / 6 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read Constant-branch translational symmetries stacked into a grid let real robots learn contact manipulation far faster by flooding the replay buffer with consistent equivalent experience.

desk verdict Real multi-seed Franka speedups from grid-scaled translational trajectory clones plus practical ego/exo buffer tricks; the invariance story is thinner than the wall-clock curves. read the letter →

arxiv 2607.26985 v1 pith:MTQPUTGW submitted 2026-07-29 cs.RO cs.AIcs.LG

Gabe Everett , Brice Gunter , Ryan Vander Stelt , Cleiver Ruiz-Martinez , Blake Hull , Juan Rojas This is my paper
classification cs.ROcs.AIcs.LG
keywords on-robotlearningsymmetryaugmentationbranchedMDPreplaybuffersuper-scalinghomographywarpingcontact-richmanipulationegocentric-exocentricvisionsample-efficientRL
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

On-robot deep reinforcement learning is still slow in wall-clock time because each physical trial is expensive. This paper argues that a simple geometric trick—treating each trajectory as the root of a constant-branch symmetry tree that tiles a planar grid of pure translations—can super-scale the amount of usable experience without new robot motion. States (proprioception plus images) are transformed in parallel; actions stay the same; rewards and dynamics are assumed invariant. Egocentric views need only pointers; exocentric views are warped at sample time with precomputed homographies. On three real contact tasks—peg insertion, cable routing, and object relocation—the method reaches high success rates in as little as roughly 11–17 minutes for the easier tasks and about 79 minutes for the harder one, beating prior on-robot baselines on speed, final success, and whole-trajectory efficiency. The claim is that super-scaling even elementary branch symmetries is enough to push practical on-robot training toward the sub-ten-minute regime for arms and humanoids.

What carries the argument

The branched MDP under constant branching: each trajectory vertex–edge pair is replicated by parallel planar translations that form a geometric grid inside a symmetry workspace; the whole grid is written into the replay buffer (with pointer-shared images or index-driven homography warps) before sampling.

What would settle it

Train the identical policy on the same peg, cable or ball task but deliberately break planar invariance—change contact friction, use a strongly non-planar object, or disable the homography and keep raw fixed-camera images—and check whether the reported wall-clock speed-ups and nAUC gains disappear or reverse.

Watch

Extended reading notes

Core claim

Modeling an episodic MDP as a symmetry tree whose constant branching produces a K-by-K grid of pure translational state transforms (actions left unchanged) populates the replay buffer with a large set of dynamics- and reward-consistent equivalents. When images are egocentric they stay invariant and are shared by pointer; when they are exocentric they are corrected by sample-time homographies. The resulting data super-scaling yields measured wall-clock convergence speed-ups of 1.37–2.17×, evaluation success-rate gains of 1.09–1.27×, fastest reported convergences of 16.6, 10.9 and 79.3 minutes on peg insertion, cable routing and object relocation, and normalized area-under-curve ratios up to 2

Load-bearing premise

Pure sideways shifts of the robot’s position, leaving the same actions and rewards, must still be valid physics for contact-rich tasks, and a flat-image warp must be a good enough stand-in for what a fixed camera would really see after that shift.

Editorial extensions

If this is right

  • On-robot contact policies for arms can reach near-ceiling success in roughly 10–20 minutes of wall-clock time on peg and cable tasks when constant-branch grids are used.
  • Replay buffers can be filled with hundreds of consistent synthetic transitions per real step without extra robot motion, raising sample efficiency while preserving real contact physics.
  • Egocentric setups need almost no extra compute; exocentric setups become usable once sample-time homographies are indexed into the buffer.
  • The same constant-branch idea is presented as a step toward sub-10-minute training on manipulation skills that transfer to humanoids.
  • Trajectory-wide nAUC, not only final success, becomes a practical way to compare whole learning curves under heavy augmentation.

Reading between the lines

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

  • If planar translational invariance is the main free lunch, adding limited rotational or reflectional branches only where contact geometry permits could push the harder multi-depth tasks without increasing artifact rate.
  • The observed drop in optimal branch count once an exocentric camera appears suggests an automatic schedule that shrinks the grid as soon as parallax or border fill begins to dominate the buffer distribution.
  • The same pointer-plus-index buffer design could be reused for other cheap invariances (lighting, gripper width) that do not require new physical trials.
  • Emergent sliding and dragging strategies on the soft ball imply the method may surface contact modes that pure demonstration cloning would miss, provided the reward classifier stays reliable under warped views.
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.

Referee Report

3 major / 6 minor

Summary. SymmGrid proposes trajectory-level data augmentation for on-robot RL by applying constant-branch planar translations that tile a K×K grid of symmetric state–action clones into the replay buffer. Egocentric images are shared via pointers; exocentric images are warped at sample time with a precomputed homography bank indexed by branch. The method is evaluated on three real Franka contact tasks (peg-insertion, cable-routing, object-relocation) against the SERL/RLPD stack, reporting wall-clock convergence speedups of roughly 1.37–2.17×, evaluation success-rate gains of 1.09–1.27× (and larger absolute gains on relocation), fastest convergences of 16.6 / 10.9 / 79.3 minutes, and nAUC ratios up to 2.59×.

Significance. If the reported speedups hold under fair comparison, the work is practically significant: multi-seed, full training-and-evaluation curves on physical contact tasks remain rare, and bringing peg/cable policies near the 10–20 minute regime would meaningfully lower the barrier to on-robot deployment. Strengths include (i) real-robot multi-seed curves rather than single best checkpoints, (ii) explicit ego- and exocentric handling with a compute-conscious sample-time homography design, (iii) trajectory-wide nAUC ratios alongside pointwise metrics, and (iv) an open project page. The contribution is primarily empirical/systems rather than a new theoretical symmetry guarantee; its value rests on whether the augmented experiences remain sufficiently consistent for contact-rich dynamics.

major comments (3)
  1. [Sec. IV; Sec. VI (state transform); Sec. VI-A (peg reward)] Sec. IV defines σ as reward- and dynamics-invariant (r′_i = R(σ_S(s_{i−1}), σ_A(a_i), σ_S(s_i)) and “rewards appear exactly as in τ”). Sec. VI then implements only x_eff ← x_eff + t_k, leaves actions, twists, wrenches, and fingers unchanged, and writes the transformed state as s[I, x_eff+t_k, θ, v, w, g]—omitting x_targ even though peg-insertion state includes x_targ and the sparse reward is a position-based norm (Sec. VI-A). If the target is not shifted identically, reward labels flip and the buffer is mislabeled; if it is shifted, that fact must be stated and the reward recomputed under the joint transform. This is load-bearing for the claim of “consistent” symmetric equivalences versus ordinary noisy augmentation.
  2. [Sec. IV (invariance claim); Sec. VI (wrench left unchanged); Sec. VII] Even with a correctly shifted target, pairing original contact wrenches/jamming (or cable deformation / ball squash) with a translated claimed pose is dynamics-invariant only under free-space planar translation. Sec. IV asserts invariance; the Discussion acknowledges exo parallax/artifacts but does not quantify how often contact events make the clones inconsistent. Without a controlled check (e.g., fraction of augmented transitions whose measured wrench would be physically incompatible with the new pose, or an ablation that zeros wrench under augmentation), the 1.37–2.17× speedups cannot be cleanly attributed to “symmetry super-scaling” rather than high-volume approximate data augmentation. Softening the invariance language and adding such a diagnostic would secure the central claim.
  3. [Sec. VI-C.1; Fig. 6; Sec. VII; Abstract] Object-relocation ablations (Sec. VI-C.1, Fig. 6) show best performance at K=3 (and narrower SymWS 0.175 m), not the 27×27 “super-scale” used for peg/cable. The Discussion correctly notes artifact/redundancy effects for exocentric multi-depth scenes, but this inverts the paper’s headline narrative that larger constant branching is what drives gains. The manuscript needs an explicit regime map: for which perception setups and contact regimes does increasing K help, and when does it hurt? Without it, “super-scaling” is task-contingent rather than a general principle, and the abstract’s uniform speedup framing overstates the method.
minor comments (6)
  1. [Sec. IV-A.1; Results] nAUC is used as a primary trajectory-wide metric but is never formally defined (normalization bounds, interpolation, whether training or evaluation curves). A short equation or appendix definition would make the 2.59× ratio reproducible.
  2. [Sec. II; References] Related work cites an anonymous/preliminary fractal report [9] and a second Frontiers piece [16] with 2026 dates; clarify relationship to the present constant-branch special case and ensure non-anonymous citation once available.
  3. [Fig. 5] Fig. 5 caption notes homography artifacts; consider marking border-fill and parallax regions explicitly so readers can judge severity without the project page.
  4. [Throughout] Typos and formatting: “replaly buffer”, “f ingers”, “branch symmetries have can have”, duplicated abstract text, and inconsistent hyphenation (ego- / egocentric). Clean pass needed.
  5. [Sec. VI (setup paragraph)] Hardware disparity vs original SERL (RTX 4070 vs 4090) is rightly noted as making results stronger; state wall-clock normalization or learner throughput so readers can separate algorithmic from compute effects.
  6. [Sec. VI-B.1] Cable-routing evaluation nAUC gain is only ~9% despite large training-curve separation; a brief remark on seed variance or early-stopping choice would help interpret Table I / Fig. 4.

Circularity Check

0 steps flagged · score 0.0 of 10

Empirical systems paper: speedups are measured against an external baseline on hardware, not derived by equating fitted inputs to outputs.

full rationale

SymmGrid’s load-bearing claims are wall-clock convergence ratios, success-rate improvements, and nAUC ratios versus SERL/RLPD on three real Franka contact tasks. Those quantities are obtained by running the same RLPD stack with and without constant-branch translational buffer augmentation and reporting timed success curves; they are not algebraic consequences of a fitted constant or of a self-cited uniqueness theorem. The symmetry operator is stated as an invariance assumption (σ reward- and dynamics-preserving planar translations, actions left unchanged, sample-time homographies for exocentric views), then tested experimentally—not proven from prior self-work and then re-labeled a prediction. Citations to the authors’ preliminary fractal/branch reports supply motivation and hyper-parameter defaults (grid size, SymWS width, buffer capacity) but do not force the measured success-rate trajectories. SERL is an independent external baseline. No step reduces Eq. X to Eq. Y by construction, and no fitted parameter is renamed a first-principles forecast. Circularity score is therefore 0; any remaining concerns (whether pure translations truly preserve contact rewards/wrenches, or whether homography artifacts corrupt the buffer) are correctness/assumption risks, not circular derivation.

Assumptions & free parameters 5 free parameters · 6 assumptions · 3 invented entities

The central speedup claim rests on standard off-policy RL machinery (MDP, replay, RLPD/SERL), the modeling choice that planar translations of trajectories are reward- and dynamics-invariant with actions held fixed, engineering choices for image handling (pointers, homographies), and several numeric knobs (K, workspace width, buffer size, demo count) selected from prior tuning rather than derived. No new physical entities are postulated; ‘branched MDP’ and ‘SymWS’ are organizational constructs for the augmentation scheme.

free parameters (5)
  • grid size K (branches per side) = 27 (primary); 9 or 3 best for exo relocation
    Number of parallel translations per axis; main experiments use K=27 for peg/cable and ablate 27/9/3 for relocation. Chosen from preliminary ‘optimal’ search, not derived.
  • symmetry workspace width (SymWS) = 0.30 m (peg/cable); 0.175 m best for relocation
    Physical extent of the translation grid around the end-effector; directly sets spacing of augmented states.
  • replay buffer capacity = 3.6 million entries
    Sized to hold super-scaled transitions; stated as 3.6M entries from preliminary work.
  • demonstration count for bootstrap = 20 (peg, cable); 30 (object relocation)
    Prior data volume for RLPD; affects early learning and is chosen by task, not derived.
  • homography calibration / edge-fill policy = M_target given; inlier RMSE 1.387 px
    M matrix from 20 point correspondences and edge-pixel replication for out-of-frame samples; quality of warped exo images depends on these choices.
assumptions (6)
  • domain assumption Standard continuous MDP and off-policy actor-critic learning with replay (RLPD/DrQ-style) yield improved policies when the buffer is enriched with valid experience.
    Entire method is an augmentation layer on SERL/RLPD; no new convergence theorem is proved (Sec. III–IV).
  • domain assumption Global planar translations σ_t of proprioceptive pose, with actions left unchanged, produce reward-preserving and dynamics-admissible alternate trajectories for the contact tasks studied.
    Stated as invariant one-to-one mappings on Γ in Sec. IV; load-bearing for treating augmented samples as valid RL experience.
  • domain assumption Egocentric wrist images are invariant under the chosen translations; exocentric images can be adequately approximated by planar homographies with edge fill.
    Sec. V; Discussion later admits parallax and artifacts limit large grids.
  • domain assumption Sparse task rewards (norm-based or image classifier threshold 0.5) and impedance control at 1 kHz match the SERL task definitions closely enough for fair comparison.
    Sec. III-C and VI; comparison validity depends on this stack parity.
  • standard math Affine/Euclidean group actions and IFS-style branching are well-defined geometric operations on poses.
    Sec. III-A uses standard E(n) / orthogonal+translation language.
  • ad hoc to paper Constant branching into a K×K grid is the right factorization of ‘branched symmetries’ for super-scaling (versus richer fractal schedules).
    Design choice emphasized in Intro/Sec. IV; justified empirically, not uniquely derived.
invented entities (3)
  • SymmGrid / branched MDP with constant-branch K×K symmetry tree independent evidence
    purpose: Organize trajectory-level translational augmentations so each time step contributes a combinatorial set of symmetric transitions to the replay buffer.
    Framing device for the augmentation scheme; not a new physical law. Independent usefulness is exactly the empirical speedups claimed.
  • Symmetry workspace (SymWS) centered on the end-effector
    purpose: Bound and space the grid translations inside the robot workspace with explicit edge spacing.
    Engineering construct defined in Sec. IV-A; parameters chosen by tuning.
  • Sample-time indexed homography bank for exocentric invariance independent evidence
    purpose: Avoid storing K² warped images by remapping the original frame on batch sample using a stored transform index.
    Systems technique combining standard homographies with buffer indexing (Sec. V).

how reviews work

0 comments
Cite this review

Pith. "Pith review of SymmGrid: Super-Scaling On-Robot Learning with Parallelized Symmetries and Egocentric-Exocentric Visual Perception." pith.science (2026). https://pith.science/paper/MTQPUTGW

@misc{pith2026260726985,
  author       = {Pith},
  title        = {Pith review of: SymmGrid: Super-Scaling On-Robot Learning with Parallelized Symmetries and Egocentric-Exocentric Visual Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MTQPUTGW}},
  note         = {Machine review of arXiv:2607.26985}
}
read the original abstract

Deep reinforcement policy learning directly in physical robots (on-robot learning) remains bottlenecked by slow wall-clock training times. We present SymmGrid, a trajectory level augmentation framework inspired by parallelized symmetries that super-scales group transformations to significantly accelerate on-robot learning in both egocentric and exocentric visual setups. We model a Markov Decision Process (MDP) under a symmetry tree, in which state-action pairs have admissible parallelized invariant transformations that yield a geometric grid structure. The state is modelled with ego- or exocentric images and proprioception information. The latter require special treatment, in the form of homographies, to warp visual scenes in line with their corresponding spatial transformations. These parallelized transformations produce a large set of unique symmetric equivalences that populate the replay buffer with diverse and consistent experiences that speed up learning and improve performance. We present extensive training and evaluations performed directly on real robot manipulation contact tasks including peg-insertions, cable routing, and object relocations. Relative to SOTA, SymmGrid achieved wall-clock training convergence speed-ups of 1.37-2.17x, evaluation success rate improvements of 1.09x-1.27x, fastest training convergence times of 16.6, 10.9, and 79.3 minutes respectively. For trajectory wide assessments, we used normalized area under the curve (nAUC) ratios. SymmGrid achieved improvements of up to 2.59x. These results confirm that simple branch symmetries can have an outsized result due to super-scaling and bring us closer to sub-10 minute on-robot learning training in manipulation tasks suitable for arms and humanoids. The project page is available at symmgrid-robot.github.io

Figures

Figures reproduced from arXiv: 2607.26985 by the authors.

Figure 1
Figure 1. Real Franka research robot examples of manipulation contact tasks. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Constant branching visualization. Iterating from one grid corner to the other: tkx,ky = SymW S(0,0) + (2 ∗ bi,j − 1) ∗ SymW S 2 ∗ bch . (1) 1) Measuring Performance: In DRL, when measuring learning speed-ups in robot learning, pointwise metrics are typically used. For example, the number of samples to achieve a particular return, or the maximum performance in the task. We believe trajectory-wide metrics are particul… view at source ↗
Figure 3
Figure 3. SymmGrid vs. SERL peg-insertion results for on-robot learning. Trained across five seeds using a constant grid size [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: SymmGrid vs. SERL cable-routing results for on-robot learning. Trained across five seeds using a constant grid size [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of homography shifts in relation to end [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Object Relocation Forward Training Results: 3 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visualization of emergent behaviors in moving a [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 5 linked inside Pith

  1. [1]

    Goldstein, C

    H. Goldstein, C. P. Poole, J. Safko, and others,Classical mechanics. Addison-wesley Reading, MA, 1950, vol. 2

    1950

  2. [2]

    Invariant Transform Experience Replay: Data Augmentation for Deep Reinforcement Learning,

    Y . Lin, J. Huang, M. Zimmer, Y . Guan, J. Rojas, and P. Weng, “Invariant Transform Experience Replay: Data Augmentation for Deep Reinforcement Learning,”IEEE Robotics and Automation Letters, vol. 5, no. 4, pp. 6615–6622, 2020. [Online]. Available: https://ieeexplore.ieee.org/document/9158366/

    arXiv 2020

  3. [3]

    Hindsight Experience Replay,

    M. Andrychowicz, F. Wolski, A. Ray, J. Schneider, R. Fong, P. Welinder, B. McGrew, J. Tobin, P. Abbeel, and W. Zaremba, “Hindsight Experience Replay,”Advances in neural information processing systems (NeurIPS), vol. 30, no. Nips, 2017. [Online]. Available: http://arxiv.org/abs/1707.01495

    arXiv 2017

  4. [4]

    Equivariant $Q$ Learning in Spatial Action Spaces,

    D. Wang, R. Walters, X. Zhu, and R. Platt, “Equivariant $Q$ Learning in Spatial Action Spaces,” inConference on Robot Learning. PMLR, 1 2021, pp. 1713–1723. [Online]. Available: http://arxiv.org/abs/2110.15443

    arXiv 2021

  5. [5]

    Reinforcement learning with augmented data,

    M. Laskin, K. Lee, A. Stooke, L. Pinto, P. Abbeel, and A. Srinivas, “Reinforcement learning with augmented data,”Advances in Neural Information Processing Systems, vol. 33, pp. 19 884–19 895, 4 2020. [Online]. Available: http: //arxiv.org/abs/2004.14990https://proceedings.neurips.cc/paper files/ paper/2020/file/e615c82aba461681ade82da2da38004a-Paper.pdf

    arXiv 2020

  6. [6]

    Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels,

    I. Kostrikov, D. Yarats, and R. Fergus, “Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels,” in International conference on learning representations (ICLR). arXiv, 4 2020. [Online]. Available: http://arxiv.org/abs/2004.13649

    arXiv 2020

  7. [7]

    SO(2)-EQUIV ARIANT REINFORCEMENT LEARNING,

    D. Wang, R. Walters, and R. Platt, “SO(2)-EQUIV ARIANT REINFORCEMENT LEARNING,” inInternational Conference on Learning Representations (ICLR), 2022. [Online]. Available: https://pointw.github.io/equi rl page/

    2022

  8. [8]

    On-Robot Learning With Equivariant Models,

    D. Wang, M. Jia, X. Zhu, R. Walters, and R. Platt, “On-Robot Learning With Equivariant Models,” inConference on Robot Learning (CoRL), 3 2022. [Online]. Available: http://arxiv.org/abs/2203.04923

    arXiv 2022

Show all 18 references
  1. [9]

    Towards Accelerating Deep Reinforcement Learning via Branched Symmetries,

    a. anynymous, “Towards Accelerating Deep Reinforcement Learning via Branched Symmetries,”Frontiers in Robotics & AI: Robot Learning and Evolution, 6 2026. [Online]. Available: https://www.frontiersin.org/ journals/robotics-and-ai/articles/10.3389/frobt.2026.1791812/abstract

    2026

  2. [10]

    Corymb Inflorescences,

    Master Gardeners of Northern Virgina, “Corymb Inflorescences,” 5 2026

    2026

  3. [11]

    Fractals and Self Similarity,

    J. E. Hutchinson, “Fractals and Self Similarity,”Mathematics Journal, vol. 30, no. 5, pp. 713–747, 1981

    1981

  4. [12]

    Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor,” pp. 1861–1870, 7 2018. [Online]. Available: https://proceedings.mlr.press/v80/haarnoja18b.html

    2018

  5. [13]

    SERL: A Software Suite for Sample-Efficient Robotic Reinforcement Learning,

    J. Luo, Z. Hu, C. Xu, S. Gadipudi, A. Sharma, R. Ahmad, S. Schaal, C. Finn, A. Gupta, and S. Levine, “SERL: A Software Suite for Sample-Efficient Robotic Reinforcement Learning,” in2024 IEEE International Conference on Robotics and Automation (ICRA), Yokohama, Japan, 2024, pp....

    2024

  6. [14]

    Efficient online reinforcement learning with offline data,

    P. J. Ball, L. Smith, I. Kostrikov, and S. Levine, “Efficient online reinforcement learning with offline data,” inInternational Conference on Machine Learning, 2023, pp. 1577–1594

    2023

  7. [15]

    Basic Concepts of the homography explained with code,

    OpenCV, “Basic Concepts of the homography explained with code,” 7 2026. [Online]. Available: https://docs.opencv.org/3.4.20/d9/dab/ tutorial homography.html

    2026

  8. [16]

    Exploring Deep Reinforcement Learning Acceleration by Superscaling Data Augmentation via Fractal Symmetries,

    R. Vander Stelt, C. Ruiz-Martinez, C. Rosen, B. Hull, and J. Rojas, “Exploring Deep Reinforcement Learning Acceleration by Superscaling Data Augmentation via Fractal Symmetries,”Frontiers in Robotics and AI - Robot Learning and Evolution, 6 2026

    2026

  9. [17]

    FMB: a Functional Manipulation Benchmark for Generalizable Robotic Learning,

    J. Luo, C. Xu, F. Liu, L. Tan, Z. Lin, J. Wu, P. Abbeel, and S. Levine, “FMB: a Functional Manipulation Benchmark for Generalizable Robotic Learning,”The International Journal of Robotics Research, vol. 44, no. 4, pp. 592–606, 9 2024. [Online]. Available: https://functional-ma...

    2024

  10. [18]

    [Online]

    RAIL-Berkeley, “SERL,” 6 2024. [Online]. Avail- able: https://github.com/rail-berkeley/serl/blob/main/docs/real franka. md#1-peg-insertion-%F0%9F%93%8D

    2024

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.