Wbc-G1-SGS — success-guided sampling

Wbc-G1-SGS — success-guided sampling#

Task id: Wbc-G1-SGS · In-tree reference on robot g1.

``Wbc-G1`` with success-guided sampling (SGS) — the same rewards, observations, terminations and assistive wrench; only where episodes start changes. Instead of the hardest clip bins, adaptive sampling favours bins the policy survives about two thirds of the time, so bins it can’t do yet (or ever: retargeting glitches) stop eating starts.

  • Preset: apply_wbc + apply_sgs (presets/sgs.py)

  • Builder: g1_wbc_sgs_env_cfg()

  • Paper: arXiv:2610.12465

What this task changes#

RSI only:

Field

Wbc-G1

Wbc-G1-SGS

strategy

similarity_ema (hardest first)

success_target

Bin signal

1 − mean reward-aligned similarity

1 if the episode terminated before its clip ended

alpha

0.005

0.01 (≈ last 100 episodes per bin)

Kernel

softmax(failure / T_eff)

target success t = 0.66, κ = 5, T = 2

How the kernel works: Adaptive sampling (RSI); every field: RSI configuration.

Train & play#

uv run wbc-mjlab-train --task Wbc-G1-SGS --dataset lafan
uv run wbc-mjlab-play --task Wbc-G1-SGS --dataset lafan --viewer viser

Logs: logs/rsl_rl/wbc_g1_sgs/<run>/

On other tasks and robots, turn on the sgs module in a Studio preset (WBC Studio), or call apply_sgs(cfg) after a task’s presets in code.

Comparing with Wbc-G1#

Train both on the same dataset, envs and seeds, then compare tracking evaluated over the whole library (uniform starts), not the training reward — the two tasks start episodes in different places, so their training curves aren’t directly comparable. The Studio’s Sampling chart shows how concentrated each sampler is (sampling_entropy, sampling_top1_prob).

Tips#

  • Lower rsi.target_success (for example 0.5) pushes harder; target_concentration=1 gives a much broader kernel (the paper’s manipulation setting).

  • Gains are expected mostly on large, mixed libraries (LAFAN and up); on a handful of clips every bin is learnable and the two samplers behave alike.