# The bottleneck in robot RL is sampling, not algorithms

The bottleneck in robot RL is sampling, not algorithms

Full article: https://haiguangboy.com/posts/pi-rl-chelsea-finn-talk

## Core approach

· The RL loop was specifically modified: humans intervene to terminate early or redirect rollouts that have already dead-ended; the value function is amortized across prompts and tasks, rather than re-estimated with each attempt.
· The control interface remains structured: the model outputs target joint positions plus 3D gripper positions, handed to a PD controller, without end-to-end raw torque. She framed this as a deliberate trade-off, not a core limitation.

## Key results

· The RL loop was specifically modified: humans intervene to terminate early or redirect rollouts that have already dead-ended; the value function is amortized across prompts and tasks, rather than re-estimated with each attempt.
· The control interface remains structured: the model outputs target joint positions plus 3D gripper positions, handed to a PD controller, without end-to-end raw torque. She framed this as a deliberate trade-off, not a core limitation.

## Comparison of related approaches

- Opposite approach · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: As reported: general robot models may already be beating specialized pipelines (π0.7 claim) contradicts joint co-training on heterogeneous human-robot data is structurally suboptimal
- Opposite approach · [Qianxun Intelligence Han Fengtao: Embodied intelligence enters the 'speed competition' phase](https://haiguangboy.com/posts/qianxun-general-brain-data-loop) `wx_ChinaWealth_20260719_2026_07`: As reported: the bottleneck for RL in robotics is physical rollout cost, not just algorithm quality contradicts core judgment: the bottleneck is not in hardware but in the brain; the robot market size is only about 1% of the automotive market
- Same approach · [Robots begin to 'stand in the light': Lingchu Intelligence enters optical module production lines](https://haiguangboy.com/posts/lingchu-optical) `wx_GalaxyFrequency_20260718_2026_07`: Source level: third-party secondhand account of a talk, single source, all figures unverified independently validates boundary: all from Lingchu's own statements plus media retelling, key figures lack third-party independent verification
- Same approach · [Liang Wenfeng's four-hour investor meeting transcript](https://haiguangboy.com/posts/liangwenfeng-world-model) `wx_elsewhere_20260722_2026_07`: Source level: third-party secondhand account of a talk, single source, all figures unverified independently validates ★boundary: secondhand compilation of a non-public meeting, media explicitly notes 'some wording may differ slightly from the original', no recording or raw records to corroborate
- Same approach · [Sudu Technology's WAIC debut: Reality is the endpoint for robots, not the boundary for training](https://haiguangboy.com/posts/sudu-third-position) `wx_LatePost_20260718_2026_07`: Source level: third-party secondhand account of a talk, single source, all figures unverified independently validates boundary: single founder interview, core tech and business claims are self-reported, 20B valuation/10K GPU cluster/POC progress unverified by third parties
- Same approach · [sunday_blog_20260717](https://haiguangboy.com/posts/sunday_blog_20260717) `sunday_blog_20260717_2026_07`: As reported: general robot models may already be beating specialized pipelines (π0.7 claim) validates route bet: strong pretraining rewrites the scaling equation—extremely small amounts of internal data drive long-tail improvements

## Boundaries

The poster was a live audience member, not from π, and the content is a secondhand account. The 700 robot-days, 500K tokens, and 10–15 minute windows should only be treated as 'alleged talk figures', not officially released π parameters.

## Author's judgment (not paper content, cross-paper synthesis)

Lining up the three judgments in the library, there are three different answers to 'where is the bottleneck': this one says physical rollout cost; another judgment says the bottleneck is not in hardware but in the brain; and one more is harsher—the implicit premise of 'world models are irrelevant to the ceiling of intelligence' is that the data pipeline is already connected; LLMs are connected so world models are a detour, robots are not connected so they are a bridge.

The three do not actually conflict. The 'brain' one addresses capability ceilings, the 'rollout cost' one addresses how to realize capabilities—they are different segments of the same path. And another item in the library, 'high-quality references let RL train with a minimal formula', exactly corroborates this reading: the main battlefield is not finding better policy gradient variants, but increasing the information yield per real trajectory.

One can return in six months to verify: if subsequent π versions still avoid end-to-end torque, it means the structured control interface is not a transitional state but a stable choice.