[Newton] Locomotion Velocity-Flat envs#3308
Merged
AntoineRichard merged 3 commits intoisaac-sim:dev/newtonfrom Sep 18, 2025
Merged
[Newton] Locomotion Velocity-Flat envs#3308AntoineRichard merged 3 commits intoisaac-sim:dev/newtonfrom
AntoineRichard merged 3 commits intoisaac-sim:dev/newtonfrom
Conversation
AntoineRichard
requested review from
Mayankm96,
jtigue-bdai,
kellyguo11 and
pascal-roth
as code owners
AntoineRichard
force-pushed
the
antoiner/dev_newton_flat_envs
branch
from
c46679b to
af803d1
Compare
hujc7
added a commit
to hujc7/IsaacLab
that referenced
this pull request
Apr 22, 2026
…verrides) Change EventsCfg.add_base_mass from additive (-5, 5) kg uniform to multiplicative (1/1.25, 1.25) log-uniform scale with distribution="log_uniform" and operation="scale". The range is symmetric in log space (geometric mean = 1.0), so the randomization has no net mass bias and is scale-invariant across robot sizes — a single default works for the 2 kg pelvis of Cassie and the 32 kg base of Anymal-D without per-robot tuning. Drop the per-robot (-1.0, 3.0) additive overrides on A1, Go1, Go2 (introduced in PR isaac-sim#3308 [Newton] Locomotion Velocity-Flat envs, 2025-09-18) which have a +1 kg mean shift that biases controllers toward always-heavier nominal — a 10% bias on A1's 10 kg base that was actively hurting convergence. ## Why the change The old (-5, 5) kg additive default was a reasonable tune for Anymal-D (≈16% range) but too wide for A1 (≈50%) and too narrow for H1 (≈10%). PR isaac-sim#3308 addressed this with per-robot (-1, 3) overrides on small quadrupeds, but the range remained asymmetric (mean shifted +1 kg) and tuning was per-robot. A multiplicative scale is more physically meaningful — mass scales in ratio with actuator torque and gravity response, so the perturbation's effect on gait is a fixed fraction regardless of robot mass. Log-uniform over [1/1.25, 1.25] gives symmetric perturbation in log space with geometric mean = 1.0, so the training distribution is unbiased. ## Ablation results (1500-iter Newton unless noted) | Robot | old default | new log-uniform | Δ | |-------------|------------:|----------------:|------------:| | A1 @300 | 3.25 | 10.00 | 3.08x | | Go1 @300 | 16.30 | 22.29 | 1.37x | | Anymal-B @300 | 10.92 | 12.47 | 1.14x | | Anymal-C @300 | 12.31 | 14.64 | 1.19x | | Go2 @1499 | 18.58 | 24.71 | 1.33x | | Anymal-D @1499 | 15.62 | 16.09 | 1.03x | A1's 3x gain is the bias-removal effect: the old (-1, 3) applied +10% mean mass shift on a 10 kg robot, forcing the controller to spend capacity compensating for a nonphysical average instead of learning the nominal gait. The new default's geometric mean 1.0 removes that bias entirely. Biped configurations (H1, Cassie) are handled in the follow-up PR isaac-sim#5298. Cassie specifically requires an asymmetric heavier-bias override due to its closed-loop Achilles coupling being mass-sensitive; that override ships alongside the biped changes.
hujc7
added a commit
to hujc7/IsaacLab
that referenced
this pull request
Apr 22, 2026
…verrides) Change EventsCfg.add_base_mass from additive (-5, 5) kg uniform to multiplicative (1/1.25, 1.25) log-uniform scale with distribution="log_uniform" and operation="scale". The range is symmetric in log space (geometric mean = 1.0), so the randomization has no net mass bias and is scale-invariant across robot sizes — a single default works for the 2 kg pelvis of Cassie and the 32 kg base of Anymal-D without per-robot tuning. Drop the per-robot (-1.0, 3.0) additive overrides on A1, Go1, Go2 (introduced in PR isaac-sim#3308 [Newton] Locomotion Velocity-Flat envs, 2025-09-18) which have a +1 kg mean shift that biases controllers toward always-heavier nominal — a 10% bias on A1's 10 kg base that was actively hurting convergence. ## Why the change The old (-5, 5) kg additive default was a reasonable tune for Anymal-D (≈16% range) but too wide for A1 (≈50%) and too narrow for H1 (≈10%). PR isaac-sim#3308 addressed this with per-robot (-1, 3) overrides on small quadrupeds, but the range remained asymmetric (mean shifted +1 kg) and tuning was per-robot. A multiplicative scale is more physically meaningful — mass scales in ratio with actuator torque and gravity response, so the perturbation's effect on gait is a fixed fraction regardless of robot mass. Log-uniform over [1/1.25, 1.25] gives symmetric perturbation in log space with geometric mean = 1.0, so the training distribution is unbiased. ## Ablation results (1500-iter Newton unless noted) | Robot | old default | new log-uniform | Δ | |-------------|------------:|----------------:|------------:| | A1 @300 | 3.25 | 10.00 | 3.08x | | Go1 @300 | 16.30 | 22.29 | 1.37x | | Anymal-B @300 | 10.92 | 12.47 | 1.14x | | Anymal-C @300 | 12.31 | 14.64 | 1.19x | | Go2 @1499 | 18.58 | 24.71 | 1.33x | | Anymal-D @1499 | 15.62 | 16.09 | 1.03x | A1's 3x gain is the bias-removal effect: the old (-1, 3) applied +10% mean mass shift on a 10 kg robot, forcing the controller to spend capacity compensating for a nonphysical average instead of learning the nominal gait. The new default's geometric mean 1.0 removes that bias entirely. Biped configurations (H1, Cassie) are handled in the follow-up PR isaac-sim#5298. Cassie specifically requires an asymmetric heavier-bias override due to its closed-loop Achilles coupling being mass-sensitive; that override ships alongside the biped changes.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Brought back all the velocity environments except Digit. They are all tuned
Type of change
Screenshots
Checklist
pre-commitchecks with./isaaclab.sh --formatconfig/extension.tomlfileCONTRIBUTORS.mdor my name already exists there